Tag selection method, device and computer equipment based on large language model

Through the automated label circle selection method based on the large language model, a circle selection strategy is generated, which solves the problem of low efficiency in human label selection and realizes an efficient and accurate label circle selection process.

CN118551769BActive Publication Date: 2024-11-12ZHEJIANG NIAOCHAO SUPPLY CHAIN MANAGEMENT CO LTD +1
View PDF 8 Cites 0 Cited by

Patent Information

Application Number
CN202411008611.0
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-07-25
Publication Date
2024-11-12
Estimated Expiration
2044-07-25

AI Technical Summary

Technical Problem

When there are many tags, manual selection of tags will significantly reduce the efficiency of tag circle selection and increase the user's complexity and difficulty.

Method used

Using the automated tag circle selection method based on the large language model, by obtaining tag circle selection information, the target large language model generates matching circle selection strategies, including circle selection labels and label values, to realize the automated tag circle selection process, and reduce user manual Selected requirements.

Benefits of technology

Improve the efficiency of label circle selection work, reduce the complexity and difficulty of users, match the data processing capabilities of large language models, and do not require users to master the huge label library, which improves the accuracy and efficiency of the generation of circle selection strategies. .

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118551769B_ABST
    Figure CN118551769B_ABST
Patent Text Reader

Abstract

The embodiments of this specification provide a tag selection method, device and computer equipment based on a large language model, which relates to the field of tag selection technology. The method includes: obtaining tag selection information, which is used to indicate the description information of the object to be selected; processing the tag selection information through a target large language model to generate a target selection strategy that matches the tag selection information, the target selection strategy includes at least one selection tag of the object to be selected and a tag value corresponding to the selection tag; tag selection is performed based on the selection tag and the tag value to obtain a target group, which includes at least one target selection object indicated by the tag selection information. The method can realize an automated tag selection process and improve the efficiency of tag selection work.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] One or more embodiments of the present specification relate to the field of tag selection technology, and more specifically, to a tag selection method, apparatus, and computer device based on a large language model in the field of artificial intelligence technology. Background Art

[0002] Tag selection means: selecting objects that meet current needs from a number of objects based on preset tags; for example, for an e-commerce platform, users with a certain type of tag or multiple types of tags are selected from all platform users in order to conduct behavioral analysis on the selected users and optimize subsequent platform marketing strategies.

[0003] In the related art, a tag library and a tag selection interface are provided, and the user selects the tag to be selected in the tag selection interface so as to perform a subsequent object selection process based on the tag; however, when there are a large number of tags, manually selecting tags will greatly reduce the efficiency of the tag selection work. Summary of the invention

[0004] One or more embodiments of the present specification provide a tag selection method, apparatus, and computer device based on a large language model, which can realize automatic tag selection and improve the efficiency of tag selection work.

[0005] On the one hand, a tag selection method based on a large language model is provided, the method comprising:

[0006] Acquire label selection information, where the label selection information is used to indicate description information of the object to be selected;

[0007] Processing the label selection information through a target large language model to generate a target selection strategy that matches the label selection information, wherein the target selection strategy includes at least one selection label of the object to be selected and a label value corresponding to the selection label;

[0008] Tag selection is performed based on the selected tag and the tag value to obtain a target group, wherein the target group includes at least one target selected object indicated by the tag selection information.

[0009] In the embodiments of this specification, an automated tag selection method is provided: the user only needs to provide descriptive information about the object to be selected, that is, the tag selection information, and the large language model can understand the selection intention in the tag selection information and automatically generate a target selection strategy, that is, automatically generate a selection tag and a tag value for the subsequent selection of the target object. The purpose of selecting the target object in one sentence can be achieved; the user does not need to manually select the tag and tag value during the selection process, which reduces the complexity and difficulty of the user's selection of the object and improves the efficiency of the tag selection work; moreover, the target selection strategy that can be used to perform the selection is generated by the large language model, and the selection tag is matched by using the data processing capability of the large language model, without the user having to master the huge number of tags in the tag library, which further reduces the difficulty of the user's selection work.

[0010] In combination with the first aspect, in some implementations of the first aspect, the processing the label circle selection information through the target large language model to generate a target circle selection strategy matching the label circle selection information includes: inputting the label circle selection information into the first large language model to perform circle selection intention recognition, and determining the target object category to which the object to be circled indicated by the label circle selection information belongs; performing knowledge recall based on the target object category to obtain at least one target circle selection knowledge, and the target circle selection knowledge is used to indicate label information and circle selection strategy information related to the object to be circled; and inputting the target circle selection knowledge and the label circle selection information into the second large language model to generate the target circle selection strategy matching the label circle selection information.

[0011] In combination with the first aspect and the above-mentioned implementation manners, in some implementation manners of the first aspect, inputting the target circle selection knowledge and the label circle selection information into a second largest language model to generate the target circle selection strategy matching the label circle selection information includes: concatenating the target circle selection knowledge and the label circle selection information to obtain target prompt information; and inputting the target prompt information into the second largest language model to obtain the target circle selection strategy output by the second largest language model.

[0012] In the embodiments of the present specification, two large language models are set, the first large language model is used to identify the circle selection intention, and the second large language model is used to generate the circle selection strategy; so as to deeply understand the user's circle selection intention from the tag circle selection information, and improve the accuracy of generating subsequent circle selection strategies; moreover, identifying the circle selection intention first can further narrow the generation scope of the circle selection strategy and further improve the generation efficiency of the circle selection strategy.

[0013] In combination with the first aspect and the above-mentioned implementation manners, in some implementation manners of the first aspect, the recalling knowledge based on the target object category to obtain at least one target circled knowledge includes: extracting keywords from the tag circled information to obtain target keywords, where the target keywords are words related to the object to be circled; and recalling knowledge based on the target keywords and the target object category to obtain at least one target circled knowledge.

[0014] In combination with the first aspect and the above-mentioned implementation manners, in some implementation manners of the first aspect, the recalling of knowledge based on the target keyword and the target object category to obtain at least one of the target circled knowledge includes: selecting candidate label knowledge under the target object category from a label knowledge base, wherein the label knowledge base stores label information under each candidate object category; matching from the candidate label knowledge based on the target keyword to obtain first circled knowledge; matching from a sample knowledge base based on the target keyword to obtain second circled knowledge, wherein the sample knowledge base stores the correspondence between historical circled information and historical circled strategies; and determining the first circled knowledge and the second circled knowledge as the recalled target circled knowledge.

[0015] In the embodiments of the present specification, in order to further improve the generation accuracy of the selection strategy, a label knowledge base and a sample knowledge base are pre-built, so that based on the target keywords and selection intentions extracted from the label selection information, similar label knowledge and historical samples can be recalled from the label knowledge base and the sample knowledge base, so as to provide richer prompt information for the selection strategy generation, so as to further improve the generation accuracy of the selection strategy.

[0016] In combination with the first aspect and the above-mentioned implementation manners, in some implementation manners of the first aspect, the processing the tag circle selection information through the target large language model to generate a target circle selection strategy that matches the tag circle selection information includes: processing the tag circle selection information through the target large language model to generate at least two candidate circle selection strategies; determining the candidate usage heat of each candidate circle selection strategy in the target circle selection scenario; and determining the candidate circle selection strategy with the highest candidate usage heat as the target circle selection strategy.

[0017] In the embodiments of the present specification, considering that multiple candidate selection strategies may be generated at the same time, the candidate selection strategies may be screened based on their candidate usage popularity, so that the target selection strategy finally determined not only meets the label selection information, but also meets the requirements of the target selection scenario, thereby avoiding outputting selection strategies that will not be used in the actual selection process, and improving the accuracy of determining the target selection strategy.

[0018] In combination with the first aspect and the above-mentioned implementation manners, in some implementation manners of the first aspect, the performing tag selection based on the selection tag and the tag value to obtain the target group includes: displaying the generated target selection strategy; if a confirmation operation on the target selection strategy is received, performing tag selection based on the target selection strategy to obtain the target group; the method further includes: if a modification operation on the target selection strategy is received, obtaining the modified target selection strategy, and performing tag selection based on the modified target selection strategy to obtain the target group.

[0019] In combination with the first aspect and the above-mentioned implementation manner, in some implementation manners of the first aspect, the method further includes: determining a target prediction loss of the target large language model based on a difference between the target selection strategy and the modified target selection strategy; and updating a model parameter of the target large language model based on the target prediction loss.

[0020] In an embodiment of the present specification, in order to avoid an invalid selection process, after the target selection strategy is generated, the target selection strategy can also be displayed so that the user can finally determine whether the subsequent selection process can be continued; if the target selection strategy is different from the user's expectations, the user can directly modify it so that the subsequent selection process can be performed according to the modified selection strategy; so that the accuracy of subsequent selection can be improved through a simple confirmation operation; in addition, after the user modifies the selection strategy, the target large language model can be re-optimized based on the difference between the modified target selection strategy and the target selection strategy generated by the model to improve the accuracy of the subsequent large language model generation selection strategy.

[0021] In combination with the first aspect and the above-mentioned implementation manners, in some implementation manners of the first aspect, the method further includes: obtaining sample circle selection information and a sample circle selection category indicated by the sample circle selection information; inputting the sample circle selection information into the first largest language model to perform circle selection intention recognition, and obtaining a predicted circle selection category output by the first largest language model; and training the first largest language model based on the difference between the predicted circle selection category and the sample circle selection category.

[0022] In the embodiments of the present specification, in order to enable the first language model to have the function of identifying circle selection intentions, the model is trained in the early stage using sample circle selection information and its corresponding sample circle selection categories, so that the first language model learns the ability to extract sample circle selection categories.

[0023] In combination with the first aspect and the above-mentioned implementation manners, in some implementation manners of the first aspect, the method further includes: obtaining sample selection information, sample selection knowledge, and a sample selection strategy, where the sample selection strategy is a selection strategy corresponding to the sample selection information and the sample selection knowledge; splicing the sample selection information and the sample selection knowledge into sample prompt information; inputting the sample prompt information into the second largest language model to obtain a predicted selection strategy output by the second largest language model; and training the second largest language model based on a difference between the predicted selection strategy and the sample selection strategy.

[0024] In the embodiments of the present specification, in order to enable the second largest language model to have the ability to generate a circle selection strategy, the model is trained in the early stage using sample circle selection information, sample circle selection knowledge and sample circle selection strategy, so that the second largest language model learns the ability to generate a circle selection strategy according to the prompt information.

[0025] In a second aspect, a label selection device based on a large language model is provided, the device comprising:

[0026] A first acquisition module is used to acquire label selection information, where the label selection information is used to indicate description information of the object to be selected;

[0027] A first generating module is used to process the label selection information through a target large language model to generate a target selection strategy that matches the label selection information, wherein the target selection strategy includes at least one selection label of the object to be selected and a label value corresponding to the selection label;

[0028] The first circle selection module is used to perform label circle selection based on the circle selection label and the label value to obtain a target group, wherein the target group includes at least one target circle selection object indicated by the label circle selection information.

[0029] In combination with the second aspect, in some implementations of the second aspect, the first generating module is further used to: input the label circle selection information into a first large language model to perform circle selection intention recognition, and determine the target object category to which the object to be circled indicated by the label circle selection information belongs; perform knowledge recall based on the target object category to obtain at least one target circle selection knowledge, and the target circle selection knowledge is used to indicate label information and circle selection strategy information related to the object to be circled; input the target circle selection knowledge and the label circle selection information into a second large language model to generate the target circle selection strategy matching the label circle selection information.

[0030] In combination with the second aspect and the above-mentioned implementation manner, in some implementation manners of the second aspect, the first generation module is further used to: splice the target circle selection knowledge and the label circle selection information to obtain target prompt information; input the target prompt information into the second largest language model to obtain the target circle selection strategy output by the second largest language model.

[0031] In combination with the second aspect and the above-mentioned implementation manner, in some implementation manners of the second aspect, the first generation module is further used to: perform keyword extraction on the tag circle selection information to obtain target keywords, where the target keywords are words related to the object to be circled; perform knowledge recall based on the target keywords and the target object category to obtain at least one of the target circled knowledge.

[0032] In combination with the second aspect and the above-mentioned implementation manners, in some implementation manners of the second aspect, the first generation module is further used to: select candidate label knowledge under the target object category from a label knowledge base, the label knowledge base storing label information under each candidate object category; match from the candidate label knowledge based on the target keyword to obtain first circled knowledge; match from a sample knowledge base based on the target keyword to obtain second circled knowledge, the sample knowledge base storing the correspondence between historical circled information and historical circled strategies; determine the first circled knowledge and the second circled knowledge as the recalled target circled knowledge.

[0033] In combination with the second aspect and the above-mentioned implementation manner, in some implementation manners of the second aspect, the first generation module is further used to: process the label selection information through the target large language model to generate at least two candidate selection strategies; determine the candidate usage heat of each candidate selection strategy in the target selection scenario; and determine the candidate selection strategy with the highest candidate usage heat as the target selection strategy.

[0034] In combination with the second aspect and the above implementation manner, in some implementation manners of the second aspect, the first selection module is used to: display the generated target selection strategy; if a confirmation operation on the target selection strategy is received, perform label selection based on the target selection strategy to obtain the target group; the device also includes: a second selection module is used to: if a modification operation on the target selection strategy is received, obtain the modified target selection strategy, and perform label selection based on the modified target selection strategy to obtain the target group.

[0035] In combination with the second aspect and the above-mentioned implementation manner, in some implementation manners of the second aspect, the device also includes: a first determination module, used to determine the target prediction loss of the target large language model based on the difference between the target selection strategy and the modified target selection strategy; an updating module, used to update the model parameters of the target large language model based on the target prediction loss.

[0036] In combination with the second aspect and the above-mentioned implementation manner, in some implementation manners of the second aspect, the device further includes: a second acquisition module, used to obtain sample circle selection information and the sample circle selection category indicated by the sample circle selection information; a second determination module, used to input the sample circle selection information into the first large language model for circle selection intention recognition, and obtain the predicted circle selection category output by the first large language model; and a first training module, used to train the first large language model based on the difference between the predicted circle selection category and the sample circle selection category.

[0037] In combination with the second aspect and the above-mentioned implementation manner, in some implementation manners of the second aspect, the device further includes: a third acquisition module, used to acquire sample selection information, sample selection knowledge and sample selection strategy, wherein the sample selection strategy is a selection strategy corresponding to the sample selection information and the sample selection knowledge; a splicing module, used to splice the sample selection information and the sample selection knowledge into sample prompt information; a third determination module, used to input the sample prompt information into the second largest language model to obtain a predicted selection strategy output by the second largest language model; and a second training module, used to train the second largest language model based on the difference between the predicted selection strategy and the sample selection strategy.

[0038] On the other hand, a computer device is provided, comprising a processor and a memory, wherein the memory stores at least one program, and the at least one program is loaded and executed by the processor to implement the tag selection method based on a large language model as described in the above aspects.

[0039] On the other hand, a computer-readable storage medium is provided, in which at least one program is stored, and the at least one program is loaded and executed by a processor to implement the tag selection method based on a large language model as described in the above aspects.

[0040] On the other hand, a computer program product comprising instructions is provided. When the computer program product runs on the computer or processor, the computer or processor executes the tag selection method based on a large language model in the above-mentioned first aspect or any possible implementation of the first aspect. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1is a schematic flow chart of a tag selection method based on a large language model provided by one or more embodiments of this specification;

[0042] Figure 2 is a schematic flow chart of another tag selection method based on a large language model provided by one or more embodiments of this specification;

[0043] Figure 3 is a schematic diagram of a tag selection process provided by one or more embodiments of this specification;

[0044] Figure 4 is a schematic flow chart of training a first language model provided by one or more embodiments of this specification;

[0045] Figure 5 is a schematic diagram of a training process of a first large language model provided by one or more embodiments of this specification;

[0046] Figure 6 is a schematic flow chart of training a first language model provided by one or more embodiments of this specification;

[0047] Figure 7 is a schematic diagram of a training process of a second largest language model provided by one or more embodiments of this specification;

[0048] Figure 8 is a schematic flow chart of another tag selection method based on a large language model provided by one or more embodiments of this specification;

[0049] Fig. 9 is a schematic flow chart of another tag selection method based on a large language model provided by one or more embodiments of this specification;

[0050] Fig.10 It is a schematic diagram of a process of generating a circle selection strategy for a large language model provided in an embodiment of the present application;

[0051] Fig.11 It is a structural schematic diagram of a tag selection device based on a large language model provided by one or more embodiments of this specification;

[0052] Fig.12 It is a structural diagram of a computer device provided by one or more embodiments of this specification. DETAILED DESCRIPTION

[0053] The following will be combined with the accompanying drawings to clearly and thoroughly describe the technical solutions in one or more embodiments of this specification. Among them, in the description of one or more embodiments of this specification, "multiple" means two or more than two. The terms "first" and "second" are used for descriptive purposes only and cannot be understood as suggesting or implying relative importance or implicitly indicating the number of technical features indicated. Therefore, the features defined as "first" and "second" may explicitly or implicitly include one or more of the features. "And / or" in the text is only a description of the association relationship of associated objects, indicating that three relationships can exist. For example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone.

[0054] In the tag selection scenario, users often need to manually select the specific tag value corresponding to the object to be selected. In a huge tag library, manually searching and selecting the required target tag is time-consuming and laborious, and also requires users to be proficient in the various tags already in the tag library. As the number of tags gradually increases, the complexity and difficulty of tag selection work increase simultaneously.

[0055] In view of the difficulty of manually selecting tags in the existing tag circle selection process, the embodiments of this specification provide an automated tag circle selection process, which can achieve the effect of circle people with one sentence. Figure 1 This is a schematic flow chart of a tag selection method based on a large language model provided by one or more embodiments of this specification. The tag selection method based on a large language model provided by the embodiments of this specification can be applied to a computer device.

[0056] For example, Figure 1 As shown, the method 100 includes:

[0057] Step 102: Acquire tag selection information, where the tag selection information is used to indicate description information of the object to be selected.

[0058] Considering that in the related art, when circling a tag, the user needs to manually select a specific tag and tag value, and when the user does not understand the tag and tag value corresponding to the object to be circled, it often leads to circling errors, increasing the difficulty of circling. In order to reduce the difficulty of circling and improve the accuracy of circling, the embodiment of this specification uses the powerful understanding ability of the large language model for natural language. The user only needs to provide descriptive information about the object to be circled, and the large language model can automatically generate a target circling strategy, that is, automatically generate circling tags and tag values ​​for subsequent circling of target objects, without the need for the user to manually select tags and tag values, reducing the complexity and difficulty of the user's circling objects.

[0059] The improved label selection process is as follows: the user inputs a description information related to the object to be selected, which may include the object features of the object to be selected, and may also include the selection method requirements when selecting the object to be selected, such as taking the intersection of two types of objects to be selected; the corresponding computer device can obtain the label selection information, and then generate a matching target selection strategy based on the label selection information.

[0060] Optionally, the label circle selection information may be obtained through text information input by the user, or may be obtained by collecting the user's voice circle selection instruction, and the voice circle selection instruction is converted into text label circle selection information.

[0061] Step 104: Process the tag selection information through the target large language model to generate a target selection strategy that matches the tag selection information. The target selection strategy includes at least one selection tag of the object to be selected and a tag value corresponding to the selection tag.

[0062] Since the label selection information does not directly include the label value corresponding to the object to be selected, the subsequent selection process still needs to match the selected object according to the selection label and the label value. In order to understand the selection strategy indicated in the label selection information input by the user, a target large language model is set in the computer device, and the label selection information is understood and processed by the target large language model to generate a target selection strategy matching the label selection information, and the target selection strategy includes at least one selection label of the object to be selected and its corresponding label value.

[0063] For example, if the input tag selection information is "I want to know the types of goods in a barbecue shop", the target selection strategy generated based on the tag selection information can be "shop-barbecue + goods-all"; corresponding to the target selection strategy, it is known that the selected tags are shop and goods, the tag value corresponding to the shop tag is barbecue, and the tag value corresponding to the goods tag is "all".

[0064] Step 106 , performing tag selection based on the tag selection tag and the tag value to obtain a target group, wherein the target group includes at least one target selected object indicated by the tag selection information.

[0065] After generating the target selection strategy, the computer device can directly perform tag selection from the object database based on the selection tag and the tag value to obtain a number of target selection objects indicated by the tag selection information and combine them into a target group.

[0066] Optionally, after the target group is generated, subsequent selection and delivery work can be performed, for example, a number of specific APP users can be identified and coupons can be issued to the APP users in the target group; optionally, an insight report corresponding to the target group can be generated based on each target selection object in the target group.

[0067] In summary, the embodiments of this specification provide an automated tag selection method: the user only needs to provide descriptive information about the object to be selected, that is, the tag selection information, and the large language model can understand the selection intention in the tag selection information and automatically generate a target selection strategy, that is, automatically generate a selection tag and a tag value for the subsequent selection of the target object. The purpose of selecting the target object in one sentence can be achieved; the user does not need to manually select the tag and tag value during the selection process, reducing the complexity and difficulty of the user's selection of the object; moreover, the target selection strategy that can be used to perform the selection is generated by the large language model, and the selection tag is matched by using the data processing capability of the large language model, without the user having to master the huge number of tags in the tag library, further reducing the difficulty of the user's selection work.

[0068] In order to improve the generation accuracy of the circle selection strategy, two large language models are set up to perform circle selection intention recognition and circle selection strategy generation respectively. After circle selection intention recognition, label knowledge recall is also introduced to further enrich the information required for circle selection strategy generation and improve the generation accuracy of the circle selection strategy.

[0069] Figure 2 This is a schematic flow chart of another tag selection method based on a large language model provided by one or more embodiments of this specification. The tag selection method based on a large language model provided by the embodiments of this specification can be applied to a computer device.

[0070] For example, Figure 2 As shown, the method 200 includes:

[0071] Step 202: Acquire tag selection information, where the tag selection information is used to indicate description information of the object to be selected.

[0072] The implementation of step 202 may refer to step 102, and this embodiment will not be described in detail here.

[0073] Step 204: input the label circle selection information into the first language model to perform circle selection intention recognition, and determine the target object category to which the to-be-circled object indicated by the label circle selection information belongs.

[0074] The target large language model includes a first large language model and a second large language model, the first large language model is used for circle selection intention recognition, and the second large language model is used for circle selection strategy generation. Exemplarily, the first large language model can adopt LLM, and the second large language model can adopt ChatGPT 3.5, LLM, etc.

[0075] Since objects of different categories may have similar labels, for example, a merchant has a barbecue label, and a barbecue product also has a barbecue label. In order to improve the accuracy of the subsequent selection strategy and avoid errors in the subsequent selection of objects, a first language model is first set up to perform natural language understanding on the input label selection information to identify the user's selection intention. The label selection information is input into the first language model to identify the selection intention, so as to output the target object category to which the object to be selected indicated by the label selection information belongs.

[0076] Specifically, multiple candidate object categories are preset, and the first language model is used to identify the probability that the object category indicated in the label selection information is a candidate of each candidate object category, and then determine the target object category to which the object to be selected indicated by the label selection information belongs based on the candidate probability. Considering the selection scenario with multiple categories of objects, the determined target object category can be one or more.

[0077] Exemplarily, the candidate object categories may include: commodities, merchants, users, etc. The candidate object categories included in different application scenarios may be different.

[0078] Optionally, when determining the target object category based on the candidate probability, a probability threshold is set, and the candidate object category with a candidate probability greater than the probability threshold is determined as the target object category; or, the candidate object category with the largest candidate probability is directly determined as the target object category.

[0079] Step 206 , performing knowledge recall based on the target object category to obtain at least one target selection knowledge, where the target selection knowledge is used to indicate label information and selection strategy information related to the object to be selected.

[0080] In order to improve the accuracy of strategy generation, a label knowledge base and a sample knowledge base are pre-built. The label knowledge base stores label information of various candidate objects, and the sample knowledge base stores historical selection information and the used label combination (label combination is the label selection strategy). By recalling knowledge from the label knowledge base and the sample knowledge base, relevant label information and relevant selection strategy information can be supplemented for subsequent strategy generation.

[0081] Specifically, the process of building a tag knowledge base can be as follows: collect tag information of each user, merchant, and product on the platform, including tag name, tag description, and tag value, summarize the tag description through LLM, extract effective information, and finally form a structured tag knowledge base. For example, a tag knowledge base in a tabular form.

[0082] The construction process of the sample knowledge base may be: collecting some historical selection groups and their corresponding selection requirement information, determining the selection requirement information as historical selection information, determining the group labels of the historical selection groups as label combinations, and storing them in the sample knowledge base in association.

[0083] Furthermore, in the label selection application process, knowledge is recalled from a pre-built knowledge base based on the target object category to obtain at least one target selection knowledge related to the generation of a target selection strategy, and the target selection knowledge is used to indicate label information and selection strategy information related to the object to be selected.

[0084] In order to improve the accuracy of recalled knowledge and reduce the recalled circled knowledge with less relevance, keyword extraction is performed on the label circled information to obtain target keywords related to the object to be circled in the label circled information, so as to combine the target keywords and the target object category to jointly recall knowledge and obtain at least one target circled knowledge.

[0085] Considering that there are two knowledge bases and the two knowledge bases store different information, the specific recall methods are also different when recalling knowledge. Correspondingly, in an exemplary example, the step of performing knowledge recall based on target keywords and target object categories to obtain at least one target circled knowledge may also include steps 206A to 206D.

[0086] Step 206A: Select candidate tag knowledge under the target object category from the tag knowledge base, where the tag knowledge base stores tag information under each candidate object category.

[0087] Since the label knowledge base stores label information of various candidate objects, in order to recall the target circled knowledge related to the label circled information, firstly, based on the target object category, the candidate label knowledge of the same category is recalled from the label knowledge base, and then the target circled knowledge obtained by this label circled information is screened.

[0088] Exemplarily, if the label knowledge base stores label information corresponding to three candidate object categories of users, commodities, and merchants, and the target object category is commodity, the label information under the commodity is first recalled based on the target object category as candidate label information.

[0089] Step 206B, matching from candidate tag knowledge based on the target keyword to obtain first circled knowledge.

[0090] The recalled candidate label knowledge is only the label knowledge under the same category, which contains a lot of label information that is irrelevant to the label selection information this time. In order to eliminate this part of irrelevant label information, matching is performed from the candidate label knowledge based on the target keyword to screen out the first selected knowledge. Specifically, when matching, a vector model can be used to represent the target keyword and the candidate label knowledge respectively, and the vector similarity between the vector of the target keyword after the representation processing and the vector of each candidate label knowledge is compared to recall the candidate label knowledge with a higher vector similarity to the target keyword, for example, recall the candidate label knowledge with a vector similarity higher than the similarity threshold as the first selected knowledge.

[0091] Step 206C: Matching is performed from the sample knowledge base based on the target keyword to obtain second selection knowledge. The sample knowledge base stores the correspondence between historical selection information and historical selection strategies.

[0092] In addition to using the target keyword to match from the tag knowledge base, it is also necessary to use the target keyword to match from the sample knowledge base to match the second circled knowledge similar to the target keyword, and the second circled knowledge is the circled strategy information related to the tag circled information. Specifically, when matching, the vector model can also be used to characterize the target keyword and each historical sample in the sample knowledge base, and by comparing the vector similarity between the vector of the target keyword after the characterization process and the vector of each historical sample, the historical samples with higher vector similarity to the target keyword are recalled, for example, the historical samples with vector similarity higher than the similarity threshold are recalled as the second circled knowledge.

[0093] Exemplarily, the historical examples are historical selection information and historical selection strategies with associated relationships stored in the example knowledge base.

[0094] Step 206D: determine the first circled knowledge and the second circled knowledge as the target circled knowledge to be recalled.

[0095] After recalling the first circle-selected knowledge (label information) from the label knowledge base and recalling the second circle-selected knowledge (historical samples, used to reflect the circle-selection strategy) from the sample knowledge base, the first circle-selected knowledge and the second circle-selected knowledge can be jointly determined as the recalled target circle-selected knowledge to provide supplementary knowledge information for the subsequent generation of the circle-selection strategy.

[0096] Step 208: input the target selection knowledge and the label selection information into the second language model to generate a target selection strategy that matches the label selection information.

[0097] A second large language model is deployed in the computer device to generate a tag selection strategy. The tag selection information and the recalled target selection knowledge can be input into the large language model, and the large language model analyzes the selection intention in the tag selection information and the selection tag information provided by the target selection knowledge to generate a target selection strategy that matches the tag selection information. The target selection strategy can be a tag combination, which includes at least one selection tag and its corresponding tag value.

[0098] In an illustrative example, step 208 may also include step 208A and step 208B.

[0099] Step 208A, combining the target circle selection knowledge and the label circle selection information to obtain target prompt information.

[0100] Step 208B: input the target prompt information into the second largest language model to obtain the target selection strategy output by the second largest language model.

[0101] Since the second largest language model is a prompt learning model, when generating a target selection strategy, the target selection knowledge and the label selection information are first concatenated to obtain the target prompt information, and then the target prompt information is input into the second largest language model to prompt the second largest language model to generate a target selection strategy related to the label selection information.

[0102] Step 210 , performing tag selection based on the tag selection tag and the tag value to obtain a target group, wherein the target group includes at least one target selected object indicated by the tag selection information.

[0103] In order to avoid selection errors, after the target selection strategy is generated, the generated target selection strategy can also be displayed so that the user can confirm whether the target selection strategy is correct. If a confirmation operation is received for the target selection strategy, the tag selection can be performed based on the selection tag and tag value in the target selection strategy to obtain the target group; conversely, if a modification operation is received for the target selection strategy, the modified target selection strategy is obtained, and the tag selection can be performed based on the modified target selection strategy to obtain the target selection.

[0104] Figure 3 1 is a schematic diagram of a tag selection process provided by one or more embodiments of this specification. Figure 3As shown, after the computer device obtains the label circle selection information 301, the label circle selection information 301 is input into the first largest language model 302 for circle selection intention recognition, and the target object category 303 of the object to be circled indicated by the label circle selection information 301 is obtained; at the same time, the target keyword 304 related to the object to be circled is extracted from the label circle selection information 301, and the label knowledge is recalled from the label knowledge base and the sample knowledge base based on the target object category 303 and the target keyword 304, and the target circle selection knowledge 305 is obtained; further, the target circle selection knowledge 305 and the label circle selection information 301 are spliced ​​to obtain the target prompt information 306, and input into the second largest language model 307, and under the prompt of the target prompt information 306, a target circle selection strategy 308 is generated; and then the object is circled according to the target circle selection strategy 308 to generate a target group 309.

[0105] In the embodiments of the present specification, two large language models are set, the first large language model is used to identify the circle selection intention, and the second large language model is used to generate the circle selection strategy; so as to deeply understand the user's circle selection intention from the tag circle selection information, and improve the accuracy of generating subsequent circle selection strategies; moreover, identifying the circle selection intention first can also further narrow the generation scope of the circle selection strategy and further improve the generation efficiency of the circle selection strategy.

[0106] In order to further improve the generation accuracy of the selection strategy, a label knowledge base and a sample knowledge base are pre-built, so that based on the target keywords and selection intentions extracted from the label selection information, similar label knowledge and historical samples can be recalled from the label knowledge base and the sample knowledge base, providing richer prompt information for the selection strategy generation, so as to further improve the generation accuracy of the selection strategy.

[0107] In order to enable the first language model to have an accurate intent recognition function during the application process of tag selection, the first language model needs to be trained in advance. This embodiment exemplifies the training process of the first language model.

[0108] Figure 4 This is a schematic flow chart of training a first language model provided by one or more embodiments of this specification. The method can be applied to a computer device.

[0109] For example, Figure 4 As shown, the method 400 includes:

[0110] Step 402: Obtain sample selection information and the sample selection category indicated by the sample selection information.

[0111] Before training the first language model, it is necessary to prepare training samples for the first language model. In order to enable the first language model to accurately identify the circle selection category indicated by the circle selection intention from the input text, multiple sample circle selection information can be pre-selected and the sample circle selection category indicated by each sample circle selection information can be manually labeled. The sample circle selection category is also the sample object category of the object to be circled. That is, the training samples of the first language model include a number of first sample pairs, each of which includes sample circle selection information and its associated sample circle selection category.

[0112] Step 404: input the sample circle selection information into the first language model to perform circle selection intention recognition, and obtain the predicted circle selection category output by the first language model.

[0113] In each round of training, the sample selection information is input into the initial first language model for selection intention recognition, and the first language model recognizes the predicted selection category indicated by the sample selection information (the predicted selection category is also the predicted object category).

[0114] Step 406: training the first language model based on the difference between the predicted circled category and the sample circled category.

[0115] If the difference between the predicted circled category and the manually labeled sample circled category is large, it means that the first language model has not accurately extracted the circled intention; if the difference between the predicted circled category and the manually labeled sample circled category is small, it means that the first language model can accurately extract the circled intention. In order to enable the first language model to learn to accurately extract the object category of the object to be circled from the sample circled information, the difference between the predicted circled category and the sample circled category is compared as a loss function to iteratively train the first language model until the value of the loss function is less than a preset value, and the training of the first language model is completed.

[0116] Figure 5 FIG. 1 is a schematic diagram of a training process of a first large language model provided by one or more embodiments of this specification. Figure 5 As shown, when training the first largest language model, first obtain a training data set of the first largest language model, the training data set includes a number of training data pairs, each training data pair includes sample selection information 501 and its corresponding sample selection category 504; in each round of training, input the sample selection information 501 into the first largest language model 502 to obtain the predicted selection category 503 output by it; then train the first largest language model 502 based on the loss value between the predicted selection category 503 and the sample selection category 504; until the loss value between the predicted selection category 503 output by the first largest language model 502 and the sample selection category 504 is less than a preset value, the training of the first largest language model 502 is completed.

[0117] In the embodiment of this specification, in order to enable the first language model to have the function of identifying circle selection intentions, the model is trained in the early stage using sample circle selection information and its corresponding sample circle selection category, so that the first language model can learn to extract the ability of sample circle selection category.

[0118] In order to enable the second largest language model to generate a matching selection strategy during the tag selection application process, the second largest language model needs to be trained in advance. This embodiment exemplifies the training process of the second largest language model.

[0119] Figure 6 This is a schematic flow chart of training a first language model provided by one or more embodiments of this specification. The method can be applied to a computer device.

[0120] For example, Figure 6 As shown, the method 600 includes:

[0121] Step 602, obtaining sample selection information, sample selection knowledge and sample selection strategy, where the sample selection strategy is a selection strategy corresponding to the sample selection information and the sample selection knowledge.

[0122] Before training the second largest language model, the training samples of the second largest language model are first required. In order to enable the second largest language model to generate a selection strategy based on the input label selection information and target selection knowledge, it is possible to pre-select and obtain a number of sample selection information and sample selection knowledge, and manually annotate the sample selection information and the sample selection knowledge corresponding to the sample selection strategy. That is, the training samples of the second largest language model include a number of second sample pairs, each of which includes sample selection information + sample selection knowledge and its associated sample selection strategy.

[0123] Step 604: combine the sample selection information and the sample selection knowledge into sample prompt information.

[0124] Since the second largest language model is a prompt model, before training the second largest language model, it is necessary to first concatenate the sample selection information and sample selection knowledge in each second sample pair to obtain sample prompt information, and then train the second largest language model based on the sample prompt information and sample selection strategy.

[0125] Step 606: Input the sample prompt information into the second largest language model to obtain the prediction circle selection strategy output by the second largest language model.

[0126] In each round of training, the sample prompt information is input into the initial second largest language model, and the initial second largest language model generates a matching prediction circle strategy under the prompt of the sample prompt information. The prediction circle selection strategy includes at least one prediction circle selection label and its corresponding prediction label value.

[0127] Step 608: training the second largest language model based on the difference between the predicted circle selection strategy and the sample circle selection strategy.

[0128] If the difference between the predicted selection strategy and the sample selection strategy manually annotated is large, it means that the selection strategy generated by the second largest language model is less accurate. If the difference between the predicted selection strategy and the sample selection strategy manually annotated is small, it means that the selection strategy generated by the second largest language model is more accurate and can be used in the actual selection application process. In order to enable the second largest language model to learn to accurately generate the selection strategy under the prompt of the second sample prompt information, the difference between the predicted selection strategy and the sample selection strategy is compared as the loss function to iteratively train the second largest language model until the value of the loss function is less than the preset value, and the training of the second largest language model is completed.

[0129] Figure 7 FIG. 1 is a schematic diagram of a training process of a second language model provided by one or more embodiments of this specification. Figure 7 As shown, before training the second largest language model, a training data set is first obtained, the training data set includes a number of training sample pairs, each training sample pair includes sample selection information 701, sample selection knowledge 702 and corresponding sample selection strategy 706; in each round of training, the sample selection information 701 and the sample selection knowledge 702 are spliced ​​into sample prompt information 703, and the sample prompt information 703 is input into the second largest language model 704 to obtain the predicted selection strategy 705 output by it; further, based on the loss value between the predicted selection strategy 705 and the sample selection strategy 706, the second largest language model 704 is trained; until the loss value between the predicted selection strategy 705 and the sample selection strategy 706 output by the second largest language model 704 is less than a preset value, the training of the second largest language model 704 is completed.

[0130] In other possible implementations, the first language model and the second language model can be trained simultaneously, that is, the sample label information is input into the first language model to obtain the predicted selection category output by the second language model; at the same time, the sample keywords are extracted from the sample label information, and the sample selection knowledge is recalled based on the predicted selection category and the sample keywords, and then the sample selection knowledge + sample selection information are spliced ​​into sample prompt information to input into the second language model to obtain the output predicted selection strategy, and then the first language model and the second language model are trained according to the loss value between the predicted selection strategy and the sample selection strategy. That is, the training samples during joint training are the sample selection information and its corresponding sample selection strategy.

[0131] In the embodiment of this specification, in order to enable the second largest language model to have the ability to generate a circle selection strategy, the model is trained in the early stage using sample circle selection information, sample circle selection knowledge and sample circle selection strategy, so that the second largest language model learns the ability to generate a circle selection strategy according to the prompt information.

[0132] The first and second language models are trained based on limited samples. Although they converged in the early training stage, there may be some special samples in the actual application process, which makes the first and second language models unable to output accurate target selection strategies. In order to continuously improve the model performance of the target language model, the model performance can be updated in time based on user feedback during the application process.

[0133] Figure 8 This is a schematic flow chart of another tag selection method based on a large language model provided by one or more embodiments of this specification. The tag selection method based on a large language model provided by the embodiments of this specification can be applied to a computer device.

[0134] For example, Figure 8 As shown, method 800 includes:

[0135] Step 802: Acquire tag circle selection information, where the tag circle selection information is used to indicate description information of the object to be circled.

[0136] Step 804: Process the tag circle selection information through the target large language model to generate a target circle selection strategy that matches the tag circle selection information.

[0137] Step 806, displaying the generated target selection strategy.

[0138] Step 808: If a modification operation of the target selection strategy is received, the modified target selection strategy is obtained, and tag selection is performed based on the modified target selection strategy to obtain a target group.

[0139] The implementation of steps 802 to 808 may refer to the above embodiments, and will not be described in detail in this embodiment.

[0140] Step 810 , determining a target prediction loss of a target large language model based on a difference between the target selection strategy and the modified target selection strategy.

[0141] Step 812: Update the model parameters of the target large language model based on the target prediction loss.

[0142] When there is a difference between the target selection strategy generated by the target large language model and the target selection strategy required by the user, the user will modify the target selection strategy, and the corresponding computer device can obtain the modified target selection strategy, and the modified target selection strategy can be used as the label selection strategy corresponding to the label selection information, and used to iteratively optimize the target large language model. Specifically, the optimization method can be: based on the difference between the target selection strategy and the modified target selection strategy, determine the target prediction loss of the target large language model, and optimize the target large language model based on the target prediction loss to update the model parameters of the target large language model.

[0143] Since the target large language model includes the first large language model and the second large language model, in one possible implementation, only the second large language model needs to be optimized, and the target prediction loss between the target circle selection strategy and the modified target circle selection strategy is used to optimize the second large language model to update the model parameters of the second large language model; in another possible implementation, the first large language model and the second large language model are optimized together, and the target prediction loss between the target circle selection strategy and the modified target circle selection strategy is used to jointly optimize the first large language model and the second large language model to update the model parameters of the first large language model and the second large language model.

[0144] Optionally, in order to further improve the optimization accuracy, when a modification operation on the target circle selection strategy is received, the circle selection intention output by the first largest language model, that is, the target object category indicated by the label circle selection information, can also be displayed, and the user determines whether the circle selection intention is accurately recognized. If the user confirms the target object category, it is determined that the first largest language model does not need to be optimized, and only the second largest language model needs to be optimized based on the target prediction loss between the target circle selection strategy and the modified target circle selection strategy, and the model parameters of the second largest language model are updated; if the user modifies the target object category, it is determined that there is also an error in the circle selection intention recognition of the first largest language model, and it is determined that there is also an error in the first largest language model and the second largest language model; at this time, the modified target object category can also be obtained, and the first largest language model can be trained and optimized based on the first prediction loss between the target object category and the modified target object category, and the second largest language model can be trained and optimized based on the second prediction loss between the target circle selection strategy and the modified target circle selection strategy, so as to update the model parameters of the first largest language model and the second largest language model.

[0145] In the embodiments of the present specification, in order to avoid invalid selection processes, after the target selection strategy is generated, the target selection strategy can also be displayed so that the user can finally determine whether the subsequent selection process can be continued; if the target selection strategy is different from the user's expectations, the user can directly modify it so that the subsequent selection process can be performed according to the modified selection strategy; so that the accuracy of subsequent selection can be improved through simple confirmation operations; in addition, after the user modifies the selection strategy, the target large language model can be re-optimized based on the difference between the modified target selection strategy and the target selection strategy generated by the model to improve the accuracy of the subsequent large language model generation selection strategy.

[0146] In other possible application scenarios, if multiple selection strategies are generated at the same time, some selection strategies may not meet the actual use requirements. For example, a certain tag combination is not frequently used in the current application scenario, or has no subsequent analysis value. In order to make the selection strategy recommended to the user more in line with the needs, after generating the selection strategy, it is necessary to filter the selection strategy according to the preset rules.

[0147] Fig. 9 This is a schematic flow chart of another tag selection method based on a large language model provided by one or more embodiments of this specification. The tag selection method based on a large language model provided by the embodiments of this specification can be applied to a computer device.

[0148] For example, Fig. 9 As shown, method 900 includes:

[0149] Step 902: Acquire tag circle selection information, where the tag circle selection information is used to indicate description information of the object to be circled.

[0150] The implementation of step 902 may refer to the above embodiment, and this embodiment will not be described in detail here.

[0151] Step 904: Process the tag selection information through the target large language model to generate at least two candidate selection strategies.

[0152] In a possible implementation, after the computer device acquires the label selection information, the first largest language model in the target large language model is used to extract the intent of the label selection information to obtain the target object category to which the object to be selected belongs; and the label selection information is subjected to keyword extraction to obtain the target keyword, so as to recall the target selection knowledge (label knowledge and selection strategy knowledge) from the label knowledge base and the sample knowledge base respectively based on the target keyword and the target object category; and then the label selection information + the target selection knowledge are concatenated to obtain the target prompt information, which is input into the second largest language model in the target large language model to generate the selection strategy, so as to obtain at least two candidate selection strategies.

[0153] Step 906: Determine the candidate usage heat of each candidate selection strategy in the target selection scenario.

[0154] When generating multiple similar candidate selection strategies, there may be selection strategies that meet the tag selection information, but are less frequently used or generally not used in the current selection scenario. In order to ensure that the target selection strategy that is finally executed is the selection strategy that will be adopted in the actual selection process, in a possible implementation, the candidate usage heat of each candidate selection strategy in the target selection scenario is also determined, and then the target selection strategy is selected according to the candidate usage heat.

[0155] Optionally, the method for determining the candidate usage heat can be: pre-constructing a heat prediction model, which is used to predict the usage heat of the input candidate selection strategies to obtain the candidate usage heat of each candidate selection strategy. The heat prediction model is constructed based on several historical selection strategies and can be used to analyze the usage heat of the candidate selection strategy in the historical selection strategies.

[0156] Step 908: determine the candidate selection strategy with the highest candidate usage popularity as the target selection strategy.

[0157] The higher the candidate usage heat, the more frequently the candidate selection strategy is used in the target selection scenario, and the more likely it is the selection strategy that will actually be adopted during the selection process; the candidate selection strategy with the highest candidate usage heat is determined as the target selection strategy.

[0158] Optionally, in other possible implementations, a policy rule may be pre-set, the policy rule defines conditional features of available selection strategies, and each candidate selection strategy is screened based on the policy rule, and the candidate selection strategy that passes the policy rule is determined as the target selection strategy.

[0159] Optionally, if the tag selection information indicates the number of strategies for generating the selection strategy, the candidate selection strategy corresponding to the number of strategies is directly determined as the target selection strategy.

[0160] Step 910 , performing tag selection based on the tag selection tag and the tag value to obtain a target group, wherein the target group includes at least one target selected object indicated by the tag selection information.

[0161] Fig.10 Schematic diagram of a process of generating a circle selection strategy for a large language model provided in an embodiment of the present application. Fig.10As shown in the figure, in the intention recognition stage, after obtaining the descriptive information (label selection information) of the input object to be selected, the selection intention is recognized based on LLM (the largest language model). If the intention is clear, the selection intention (target object category) can be identified as at least one of crowd selection, product selection, merchant selection and analytical insight. If the intention is not clear, a prompt message is displayed to ask the user about the selection intention; in the label retrieval stage, the target keywords are extracted from the description information, and M3E is used to recall similar tags and historical samples from the pre-built knowledge base based on the target keywords and selection intention to obtain the target selection knowledge; among them, if similar tags can be recalled, the subsequent generation of label combination stage will continue to be executed. If similar tags are not recalled, If the target tag is not found, the tag requirements (i.e., tag selection information) are collected again; in the tag combination stage, the description information, the recalled similar tags and the historical samples are spliced ​​to construct the target prompt information, and the target prompt information is input into the LLM (the second largest language model) to generate the tag combination (i.e., the target selection strategy); in the group generation stage, first determine whether the target selection strategy meets the preset rules. If the rule requirements are met, the group is generated based on the target selection strategy; if the rule requirements are not met, the target selection strategy is tuned (or modified) and the group is generated based on the tuned or modified selection strategy; in the automated insight stage, the group report is generated by analyzing each object in the group; optionally, the group report can be displayed in a visual icon. Specifically, in the knowledge base construction stage, the original tag data (or information) of users, merchants, and products, such as tag names, tag descriptions, and tag values, can be obtained and processed into a structured tag library; the historical selection strategy and the corresponding tag combination can also be obtained and associated as historical samples to obtain a structured sample library.

[0162] The implementation of step 910 may refer to the above embodiment, and this embodiment will not be described in detail here.

[0163] In the embodiments of the present specification, considering that multiple candidate selection strategies may be generated at the same time, the candidate selection strategies may also be screened based on their candidate usage popularity, so that the target selection strategy finally determined not only meets the label selection information, but also meets the requirements of the target selection scenario, avoids outputting selection strategies that will not be used in the actual selection process, and improves the accuracy of determining the target selection strategy.

[0164] Fig.11 This is a schematic diagram of a structure of a tag selection device based on a large language model provided by one or more embodiments of this specification. The device is applied in a computer device.

[0165] For example, Fig.11 The device 1100 includes:

[0166] A first acquisition module 1102 is used to acquire tag circle selection information, where the tag circle selection information is used to indicate description information of the object to be circled;

[0167] A first generating module 1104 is configured to process the label selection information through a target large language model to generate a target selection strategy that matches the label selection information, wherein the target selection strategy includes at least one selection label of the object to be selected and a label value corresponding to the selection label;

[0168] The first circle selection module 1106 is configured to perform a tag circle selection based on the tag selection tag and the tag value to obtain a target group, wherein the target group includes at least one target circle selection object indicated by the tag circle selection information.

[0169] Optionally, the first generating module 1104 is further configured to:

[0170] Inputting the label circle selection information into the first language model to perform circle selection intention recognition, and determining the target object category to which the to-be-circled object indicated by the label circle selection information belongs;

[0171] Recalling knowledge based on the target object category to obtain at least one target selection knowledge, where the target selection knowledge is used to indicate label information and selection strategy information related to the object to be selected;

[0172] The target selection knowledge and the label selection information are input into a second language model to generate the target selection strategy that matches the label selection information.

[0173] Optionally, the first generating module 1104 is further configured to:

[0174] The target circle selection knowledge and the label circle selection information are combined to obtain target prompt information;

[0175] The target prompt information is input into the second largest language model to obtain the target selection strategy output by the second largest language model.

[0176] Optionally, the first generating module 1104 is further configured to:

[0177] Perform keyword extraction on the tag circle selection information to obtain target keywords, where the target keywords are words related to the object to be circled;

[0178] Knowledge recall is performed based on the target keyword and the target object category to obtain at least one target circled knowledge.

[0179] Optionally, the first generating module 1104 is further configured to:

[0180] Selecting candidate label knowledge under the target object category from a label knowledge base, wherein the label knowledge base stores label information under each candidate object category;

[0181] Matching the candidate tag knowledge based on the target keyword to obtain first circled knowledge;

[0182] Based on the target keyword, matching is performed from a sample knowledge base to obtain second selection knowledge, wherein the sample knowledge base stores a correspondence between historical selection information and historical selection strategies;

[0183] The first circled knowledge and the second circled knowledge are determined as the recalled target circled knowledge.

[0184] Optionally, the first generating module 1104 is further configured to:

[0185] Processing the tag selection information by using the target large language model to generate at least two candidate selection strategies;

[0186] Determine the candidate usage popularity of each candidate selection strategy in the target selection scenario;

[0187] The candidate selection strategy with the highest candidate usage popularity is determined as the target selection strategy.

[0188] Optionally, the first selection module 1106 is used to:

[0189] Displaying the generated target selection strategy;

[0190] If a confirmation operation on the target selection strategy is received, label selection is performed based on the target selection strategy to obtain the target group;

[0191] The device also includes:

[0192] The second selection module is configured to, if a modification operation on the target selection strategy is received, obtain the modified target selection strategy, and perform tag selection based on the modified target selection strategy to obtain the target group.

[0193] Optionally, the device further comprises:

[0194] A first determination module, configured to determine a target prediction loss of the target large language model based on a difference between the target selection strategy and the modified target selection strategy;

[0195] An updating module is used to update model parameters of the target large language model based on the target prediction loss.

[0196] Optionally, the device further comprises:

[0197] A second acquisition module, used to acquire sample selection information and a sample selection category indicated by the sample selection information;

[0198] A second determination module is used to input the sample circle selection information into the first language model to perform circle selection intention recognition, and obtain a predicted circle selection category output by the first language model;

[0199] The first training module is used to train the first language model based on the difference between the predicted circled category and the sample circled category.

[0200] Optionally, the device further comprises:

[0201] A third acquisition module is used to acquire sample selection information, sample selection knowledge and sample selection strategy, wherein the sample selection strategy is a selection strategy corresponding to the sample selection information and the sample selection knowledge;

[0202] A splicing module, used for splicing the sample selection information and the sample selection knowledge into sample prompt information;

[0203] A third determination module, configured to input the sample prompt information into the second largest language model to obtain a prediction circle selection strategy output by the second largest language model;

[0204] The second training module is used to train the second largest language model based on the difference between the prediction selection strategy and the sample selection strategy.

[0205] The embodiment of this specification provides a computer device, which includes a processor and a memory, wherein at least one program is stored in the memory, and the at least one program is loaded and executed by the processor to implement the tag selection method based on a large language model as provided in the above optional implementation. Optionally, the computer device can be a tag selection platform.

[0206] Fig.12 It is a structural diagram of a computer device provided by one or more embodiments of this specification.

[0207] For example, Fig.12 As shown, the computer device 1200 includes: a memory 1201, a processor 1202, and a computer program 1203 stored in the memory 1201 and running on the processor 1202, wherein when the processor 1202 executes the computer program 1203, the computer device can execute any one of the label selection methods based on the large language model introduced above.

[0208] One or more embodiments of this specification may divide the functional modules of the computer device according to the above method examples. For example, each functional module may be corresponding, or two or more functions may be integrated into one processing module. The above integrated module may be implemented in the form of hardware. It should be noted that the division of modules in one or more embodiments of this specification is schematic and is only a logical function division. There may be other division methods in actual implementation.

[0209] In the case of dividing each functional module according to each function, the computer device may include: a first acquisition module, a first generation module, a first selection module, etc. It should be noted that all relevant contents of each step involved in the above method embodiment can be referred to the functional description of the corresponding functional module, which will not be repeated here.

[0210] The computer device provided in one or more embodiments of this specification is used to execute the above-mentioned tag selection method based on a large language model, so it can achieve the same effect as the above-mentioned implementation method.

[0211] In the case of an integrated unit, the computer device may include a processing module and a storage module. The processing module may be used to control and manage the actions of the computer device. The storage module may be used to support the computer device in executing mutual program codes and data.

[0212] The processing module may be a processor or a controller, which may implement or execute various exemplary logic blocks, modules and circuits disclosed in combination with one or more embodiments of this specification. The processor may also be a combination that implements computing functions, such as a combination of one or more microprocessors, a combination of digital signal processing (DSP) and a microprocessor, etc. The storage module may be a memory.

[0213] The computer device provided in one or more embodiments of the present specification may specifically be a chip, a component or a module, and the computer device may include a connected processor and a memory; wherein the memory is used to store instructions, and when the computer device is running, the processor may call and execute the instructions so that the chip executes any one of the label selection methods based on the large language model described above.

[0214] One or more embodiments of the present specification provide a computer-readable storage medium, which stores instructions. When the instructions are executed on a computer or a processor, the computer or the processor executes any one of the label selection methods based on a large language model described above.

[0215] One or more embodiments of the present specification also provide a computer program product comprising instructions, which, when executed on a computer or a processor, enables the computer or the processor to execute the above-mentioned related steps to implement any of the label selection methods based on a large language model described above.

[0216] Among them, the computer device, computer-readable storage medium, computer program product or chip containing instructions provided in one or more embodiments of this specification are used to execute the corresponding label selection method based on the large language model provided above. Therefore, the beneficial effects that can be achieved can refer to the beneficial effects in the corresponding method provided above, and will not be repeated here.

[0217] Through the description of the above implementation methods, technical personnel in the relevant field can understand that for the convenience and simplicity of description, only the division of the above-mentioned functional modules is used as an example. In actual applications, the above-mentioned functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above.

[0218] In one or more embodiments of the present specification, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic, for example, the division of modules or units is only a logical function division, and there may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another device, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0219] The above is a description of a specific embodiment of the specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recorded in the claims can be performed in an order different from that in the embodiments and still achieve the desired results. In addition, the processes depicted in the drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0220] The above contents are only one or more specific implementations in this specification, but the protection scope of one or more embodiments of this specification is not limited thereto. Any technician familiar with the technical field can easily think of changes or replacements within the technical scope disclosed in one or more embodiments of this specification, which should be included in the protection scope of one or more embodiments of this specification. Therefore, the protection scope of one or more embodiments of this specification shall be based on the protection scope of the claims.

Claims

1. A tag selection method based on a large language model, characterized in that: The method comprises: Acquire label selection information, where the label selection information is used to indicate description information of the object to be selected; Processing the label selection information through a target large language model to generate a target selection strategy that matches the label selection information, wherein the target selection strategy includes at least one selection label of the object to be selected and a label value corresponding to the selection label; Performing tag selection based on the tag selection tag and the tag value to obtain a target group, wherein the target group includes at least one target selected object indicated by the tag selection information; The step of processing the tag circle selection information by using a target large language model to generate a target circle selection strategy matching the tag circle selection information includes: Inputting the label circle selection information into the first language model to perform circle selection intention recognition, and determining the target object category to which the to-be-circled object indicated by the label circle selection information belongs; Recalling knowledge based on the target object category to obtain at least one target selection knowledge, where the target selection knowledge is used to indicate label information and selection strategy information related to the object to be selected; Inputting the target selection knowledge and the label selection information into a second language model to generate the target selection strategy matching the label selection information; The step of performing knowledge recall based on the target object category to obtain at least one target circled knowledge includes: Perform keyword extraction on the tag circle selection information to obtain target keywords, where the target keywords are words related to the object to be circled; Based on the target keyword and the target object category, matching is performed from a label knowledge base and a sample knowledge base to obtain at least one of the target selection knowledge, wherein the label knowledge base stores label information under each candidate object category, and the sample knowledge base stores the correspondence between historical selection information and historical selection strategies.

2. The method according to claim 1, characterized in that The step of inputting the target selection knowledge and the label selection information into a second language model to generate the target selection strategy matching the label selection information includes: The target circle selection knowledge and the label circle selection information are combined to obtain target prompt information; The target prompt information is input into the second largest language model to obtain the target selection strategy output by the second largest language model.

3. The method according to claim 1, characterized in that The step of matching the target keyword and the target object category from a tag knowledge base and a sample knowledge base to obtain at least one target selection knowledge includes: Selecting candidate label knowledge under the target object category from the label knowledge base; Matching the candidate tag knowledge based on the target keyword to obtain first circled knowledge; Matching from the sample knowledge base based on the target keyword to obtain second circled knowledge; The first circled knowledge and the second circled knowledge are determined as the recalled target circled knowledge.

4. The method according to any one of claims 1 to 3, characterized in that: The step of processing the tag circle selection information by using a target large language model to generate a target circle selection strategy matching the tag circle selection information includes: Processing the tag selection information by using the target large language model to generate at least two candidate selection strategies; Determine the candidate usage popularity of each candidate selection strategy in the target selection scenario; The candidate selection strategy with the highest candidate usage popularity is determined as the target selection strategy.

5. The method according to any one of claims 1 to 3, characterized in that: The performing tag selection based on the selected tag and the tag value to obtain a target group includes: Displaying the generated target selection strategy; If a confirmation operation on the target selection strategy is received, label selection is performed based on the target selection strategy to obtain the target group; The method further comprises: If a modification operation of the target selection strategy is received, the modified target selection strategy is obtained, and tag selection is performed based on the modified target selection strategy to obtain the target group.

6. The method according to claim 5, characterized in that The method further comprises: determining a target prediction loss of the target large language model based on a difference between the target circle selection strategy and the modified target circle selection strategy; Based on the target prediction loss, model parameters of the target large language model are updated.

7. The method according to any one of claims 1 to 3, characterized in that: The method further comprises: Acquire sample selection information and the sample selection category indicated by the sample selection information; Inputting the sample circle selection information into the first language model to perform circle selection intention recognition, and obtaining a predicted circle selection category output by the first language model; The first language model is trained based on the difference between the predicted circled category and the sample circled category.

8. The method according to any one of claims 1 to 3, characterized in that: The method further comprises: Acquire sample selection information, sample selection knowledge and a sample selection strategy, wherein the sample selection strategy is a selection strategy corresponding to the sample selection information and the sample selection knowledge; splicing the sample selection information and the sample selection knowledge into sample prompt information; Inputting the sample prompt information into the second largest language model to obtain a prediction circle selection strategy output by the second largest language model; The second largest language model is trained based on the difference between the prediction circle selection strategy and the sample circle selection strategy.

9. A label selection device based on a large language model, characterized in that: The device comprises: A first acquisition module is used to acquire label selection information, where the label selection information is used to indicate description information of the object to be selected; A first generating module is used to process the label selection information through a target large language model to generate a target selection strategy that matches the label selection information, wherein the target selection strategy includes at least one selection label of the object to be selected and a label value corresponding to the selection label; A first selection module, configured to perform tag selection based on the selection tag and the tag value to obtain a target group, wherein the target group includes at least one target selection object indicated by the tag selection information; The first generating module is specifically used for: Inputting the label circle selection information into the first language model to perform circle selection intention recognition, and determining the target object category to which the to-be-circled object indicated by the label circle selection information belongs; Recalling knowledge based on the target object category to obtain at least one target selection knowledge, where the target selection knowledge is used to indicate label information and selection strategy information related to the object to be selected; Inputting the target selection knowledge and the label selection information into a second language model to generate the target selection strategy matching the label selection information; The first generating module is further specifically used for: Perform keyword extraction on the tag circle selection information to obtain target keywords, where the target keywords are words related to the object to be circled; Based on the target keyword and the target object category, matching is performed from a label knowledge base and a sample knowledge base to obtain at least one of the target selection knowledge, wherein the label knowledge base stores label information under each candidate object category, and the sample knowledge base stores the correspondence between historical selection information and historical selection strategies.

10. A computer device, comprising a processor and a memory, wherein the memory stores at least one program, and the at least one program is loaded and executed by the processor to implement the tag selection method based on a large language model as described in any one of claims 1 to 8.

11. A computer-readable storage medium, wherein at least one program is stored in the computer-readable storage medium, and the at least one program is loaded and executed by a processor to implement the tag selection method based on a large language model as described in any one of claims 1 to 8.

12. A computer program product comprising instructions, which, when executed on a computer device or a processor, enables the computer device or the processor to execute the tag selection method based on a large language model as claimed in any one of claims 1 to 8.

Citation Information

Patent Citations

  • Label associated content method and device

    CN110489649A

  • Target crowd circling method and device, equipment and storage medium

    CN114022188A

  • Data processing method and device, equipment and medium

    CN115759100A

  • Target crowd circling method and device, computer equipment and readable storage medium

    CN116483987A

  • Information screening method and device, electronic equipment, storage medium and program product

    CN118132614A