Cboth generation method and multi-modal information generation method

Abstract

The embodiment of the invention provides a copywriting generation method and a multi-modal information generation method. The copywriting generation method comprises the following steps: determining a target object and a user group associated with the target object; constructing a multi-dimensional copywriting set for the target object, selecting a matched copywriting matched with the user group in the multi-dimensional copywriting set, and determining a reference copywriting in the multi-dimensional copywriting set; constructing a copywriting generation sample pair according to the matched copywriting, and constructing a copywriting preference sample pair according to the matched copywriting and the reference copywriting; and optimizing a large language model based on the copywriting generation sample pair and the copywriting preference sample pair, and generating a target copywriting corresponding to the target object by using the optimized large language model.

Description

Technical Field

[0001] The embodiments of this specification relate to the technical field of information push, and particularly to a copywriting generation method and a multi-modal information generation method. Background Art

[0002] With the development of computer and Internet technologies, in service items that require the push of copywriting, long text generation can reflect better personalized effects. By increasing the text length, it is possible to ensure that the generated copywriting has rich details and at the same time reflect the personalized characteristics of users. However, in the field of message push, especially in the scenario of short titles, there are many constraints in personalized copywriting generation. Firstly, it is difficult for short copywriting to reflect the personalized needs and preferences of users, and the differences between different users cannot be significantly reflected. Secondly, when the category information of the push object is rigidly combined with user attributes, it may lead to a rigid expression and violate language logic. To solve this technical problem, most of the existing technologies directly use large language models to generate short copywriting. Although the copywriting can be associated with the push object, the user attributes are not fully combined in this process, resulting in a not-so-good effect of the pushed copywriting. Therefore, an effective solution is urgently needed to solve the above problems. Summary of the Invention

[0003] In view of this, the embodiments of this specification provide a copywriting generation method. One or more embodiments of this specification are also related to a multi-modal information generation method, a copywriting generation device, a multi-modal information generation device, a computing device, a computer-readable storage medium, and a computer program product to solve the technical defects existing in the prior art.

[0004] According to the first aspect of the embodiments of this specification, a copywriting generation method is provided, including: Determine a target object and the user group associated with the target object; Construct a multi-dimensional copywriting set for the target object, select a matching copywriting that matches the user group from the multi-dimensional copywriting set, and determine a reference copywriting in the multi-dimensional copywriting set; Construct a copywriting generation sample pair according to the matching copywriting, and construct a copywriting preference sample pair according to the matching copywriting and the reference copywriting; Optimize a large language model based on the copywriting generation sample pair and the copywriting preference sample pair, and use the optimized large language model to generate a target copywriting corresponding to the target object.

[0005] According to the second aspect of the embodiments of this specification, another copywriting generation method is provided, including: Determine a target commodity and the user group associated with the target commodity; Construct a multi-dimensional push copy set for the target product, select a matching copy that matches the user group from the multi-dimensional push copy set, and determine a reference copy in the multi-dimensional push copy set; Construct a copy generation sample pair according to the matching copy, and construct a copy preference sample pair according to the matching copy and the reference copy; Optimize the large language model based on the copy generation sample pair and the copy preference sample pair, and use the optimized large language model to generate the target push copy corresponding to the target product; Send the target push copy to the terminal devices of the users included in the user group.

[0006] According to the third aspect of the embodiments of the present specification, a multi-modal information generation method is provided, including: Determine the target object and the user group associated with the target object; Construct a multi-dimensional multi-modal information set for the target object, select a matching multi-modal information that matches the user group from the multi-dimensional multi-modal information set, and determine a reference multi-modal information in the multi-dimensional multi-modal information set; Construct a multi-modal information generation sample pair according to the matching multi-modal information, and construct a multi-modal information preference sample pair according to the matching multi-modal information and the reference multi-modal information; Optimize the large language model based on the multi-modal information generation sample pair and the multi-modal information preference sample pair, and use the optimized large language model to generate the target multi-modal information corresponding to the target object.

[0007] According to the fourth aspect of the embodiments of the present specification, a copy generation device is provided, including: A determination module configured to determine the target object and the user group associated with the target object; A selection module configured to construct a multi-dimensional copy set for the target object, select a matching copy that matches the user group from the multi-dimensional copy set, and determine a reference copy in the multi-dimensional copy set; A construction module configured to construct a copy generation sample pair according to the matching copy, and construct a copy preference sample pair according to the matching copy and the reference copy; A generation module configured to optimize the large language model based on the copy generation sample pair and the copy preference sample pair, and use the optimized large language model to generate the target copy corresponding to the target object.

[0008] According to the fifth aspect of the embodiments of the present specification, another copy generation device is provided, including: A commodity determination module, configured to determine a target commodity and the user group associated with the target commodity; A copywriting selection module, configured to construct a multi-dimensional push copywriting set for the target commodity, select a matching copywriting that matches the user group from the multi-dimensional push copywriting set, and determine a reference copywriting from the multi-dimensional push copywriting set; A sample construction module, configured to construct a copywriting generation sample pair according to the matching copywriting, and construct a copywriting preference sample pair according to the matching copywriting and the reference copywriting; A copywriting generation module, configured to optimize a large language model based on the copywriting generation sample pair and the copywriting preference sample pair, and use the optimized large language model to generate a target push copywriting corresponding to the target commodity; A push copywriting module, configured to send the target push copywriting to the terminal devices of the users included in the user group.

[0009] According to the sixth aspect of the embodiments of the present specification, a multi-modal information generation device is provided, including: A group determination module, configured to determine a target object and the user group associated with the target object; An information selection module, configured to construct a multi-dimensional multi-modal information set for the target object, select a matching multi-modal information that matches the user group from the multi-dimensional multi-modal information set, and determine a reference multi-modal information from the multi-dimensional multi-modal information set; A sample construction module, configured to construct a multi-modal information generation sample pair according to the matching multi-modal information, and construct a multi-modal information preference sample pair according to the matching multi-modal information and the reference multi-modal information; An information generation module, configured to optimize a large language model based on the multi-modal information generation sample pair and the multi-modal information preference sample pair, and use the optimized large language model to generate a target multi-modal information corresponding to the target object.

[0010] According to the seventh aspect of the embodiments of the present specification, a computing device is provided, including: A memory and a processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the computer-executable instructions are executed by the processor, the steps of the above copywriting generation method or multi-modal information generation method are implemented.

[0011] According to the eighth aspect of the embodiments of the present specification, a computer-readable storage medium is provided, which stores computer-executable instructions. When the instructions are executed by a processor, the steps of the above copywriting generation method or multi-modal information generation method are implemented.

[0012] According to a ninth aspect of the embodiments of the present specification, a computer program product is provided, including a computer program or instruction, which when executed by a processor, implements the steps of the above-mentioned copywriting generation method or multimodal information generation method.

[0013] For the copywriting generation method provided in this embodiment, in order to ensure that the generated copywriting better matches the user group and has a better effect, the target object and the user group associated with the target object can be determined first, and then a multi-dimensional copywriting set can be constructed for the target object first to ensure the richness of the copywriting through the multi-dimensional copywriting set. At this time, the matching copywriting that matches the user group can be selected from the multi-dimensional copywriting set, so as to ensure that the subsequent processed copywriting is associated with the user group. At the same time, the reference copywriting can be determined from the multi-dimensional copywriting set to ensure that the model can distinguish the preferences of the user group. On this basis, a copywriting generation sample pair can be constructed according to the matching copywriting, and a copywriting preference sample pair can be constructed according to the matching copywriting and the reference copywriting, and the large language model can be optimized based on the copywriting generation sample pair and the copywriting preference sample pair, so that the large language model has the ability to generate copywriting and can learn the preference knowledge of the user group. Finally, the optimized large language model can be used to generate the target copywriting corresponding to the target object. Furthermore, it is ensured that when constructing copywriting for the user group, the user attributes and the generation ability of the large language model can be fully combined, so that the target copywriting better matches the user group, thereby improving the copywriting push effect of downstream services. Description of the Drawings

[0014] Figure 1 is a schematic diagram of a copywriting generation method provided by an embodiment of the present specification; Figure 2 is a flowchart of a copywriting generation method provided by an embodiment of the present specification; Figure 3 is a flowchart of another copywriting generation method provided by an embodiment of the present specification; Figure 4 is a flowchart of a multimodal information generation method provided by an embodiment of the present specification; Figure 5 is a processing process flowchart of a copywriting generation method provided by an embodiment of the present specification; Figure 6 is a schematic structural diagram of a copywriting generation device provided by an embodiment of the present specification; Figure 7 is a schematic structural diagram of another copywriting generation device provided by an embodiment of the present specification; Figure 8 is a schematic structural diagram of a multimodal information generation device provided by an embodiment of the present specification; Figure 9It is a structural block diagram of a computing device provided by an embodiment of this specification. Detailed implementation manners

[0015] In the following description, numerous specific details are set forth in order to provide a thorough understanding of this specification. However, this specification can be implemented in many other ways different from those described herein, and those skilled in the art can make similar generalizations without departing from the connotation of this specification. Therefore, this specification is not limited by the specific implementations disclosed below.

[0016] The terms used in one or more embodiments of this specification are for the purpose of describing specific embodiments only and are not intended to limit one or more embodiments of this specification. The singular forms "a", "the", and "said" used in one or more embodiments of this specification and the appended claims are also intended to include the plural forms unless the context clearly dictates otherwise. It should also be understood that the term "and / or" used in one or more embodiments of this specification refers to and encompasses any and all possible combinations of one or more of the associated listed items.

[0017] It should be understood that although the terms first, second, etc. may be used in one or more embodiments of this specification to describe various information, such information should not be limited to these terms. These terms are only used to distinguish the same type of information from each other. For example, without departing from the scope of one or more embodiments of this specification, the first may also be referred to as the second, and similarly, the second may also be referred to as the first. Depending on the context, the word "if" as used herein may be interpreted as "when" or "while" or "in response to determining".

[0018] In addition, it should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in one or more embodiments of this specification are all information and data that have been authorized by the user or fully authorized by all parties, and the collection, use, and processing of the relevant data need to comply with the relevant laws, regulations, and standards of the relevant countries and regions, and corresponding operation entrances are provided for the user to select to authorize or reject.

[0019] In one or more embodiments of this specification, a large model refers to a deep learning model with a large number of model parameters, usually including hundreds of millions, tens of billions, hundreds of billions, trillions or even more than one quadrillion model parameters. A large model can also be called a Foundation Model. Through the pre-training of a large model with a large amount of unlabeled corpus, a pre-trained model with more than hundreds of millions of parameters is produced. This model can adapt to a wide range of downstream tasks and has good generalization ability. For example, large language models (LLMs), multi-modal pre-training models, etc.

[0020] When a large model is actually applied, only a small number of samples are needed to fine-tune the pre-trained model for application to different tasks. Large models can be widely applied in fields such as natural language processing (NLP) and computer vision. Specifically, they can be applied to tasks in the field of computer vision such as visual question answering (VQA), image captioning (IC), image generation, etc., and tasks in the field of natural language processing such as text-based sentiment classification, text summary generation, machine translation, etc. The main application scenarios of large models include digital assistants, intelligent robots, search, online education, office software, e-commerce, intelligent design, etc.

[0021] First, the noun terms involved in one or more embodiments of this specification are explained.

[0022] Offline Naive Bandit Algorithm: It is a simplified form of the Bandit algorithm, mainly used to solve decision-making problems in an uncertain environment, especially to balance the exploration and exploitation dilemmas in recommendation systems. The core idea of the Naive Bandit algorithm is: among multiple options (or called "arms"), one is selected for trial through a certain strategy, and the evaluation of each option is updated according to the result of the trial to make better choices in the future. Specifically, the Naive Bandit algorithm will first randomly experiment several times to calculate the average reward of each arm. Then, in subsequent selections, with a certain probability (such as 1 - epsilon), the arm with a higher current reward is selected (i.e., exploitation), and with a certain probability (such as epsilon), a random arm is selected for exploration. Here, epsilon is a parameter that controls the balance between exploration and exploitation.

[0023] LLM, namely Large Language Model, is a deep learning algorithm capable of performing various natural language processing (NLP) tasks. Underlying the LLM are multiple transformer models, which are a set of neural networks consisting of an encoder and a decoder with self-attention capabilities. The encoder and decoder extract meaning from a series of texts and understand the relationships between words and phrases within them. By training the transformer, it can learn to understand basic grammar, language, and knowledge.

[0024] Prompt Engineering, namely prompt engineering, is a technique for guiding a large language model (LLM) to generate specific types of outputs by designing carefully constructed prompts or inputs. The principle behind this technique is to utilize the model's sensitivity to the input. By providing prompts in a specific format or with specific content, it influences the model's internal state, thereby guiding the model to generate outputs that meet expectations.

[0025] EE: Exploitation&Exploration. In the decision-making process, the balance between EE (Exploitation and Exploration) is a key issue. Exploitation refers to choosing the currently seemingly optimal strategy based on existing experience to maximize the current gain or performance. Exploration, on the other hand, refers to trying some new strategies or methods, which may not be optimal currently but have the potential to bring higher gains or discover better solutions in the future.

[0026] In this specification, a copywriting generation method is provided. One or more embodiments of this specification are simultaneously related to a multi-modal information generation method, a copywriting generation device, a multi-modal information generation device, a computing device, a computer-readable storage medium, and a computer program product, which will be described in detail one by one in the following embodiments.

[0027] See Figure 1In the schematic diagram shown, for the copywriting generation method provided in this embodiment, in order to ensure that the generated copywriting better matches the user group and has a better effect, the target object and the user group associated with the target object can be determined first. Then, a multi-dimensional copywriting set can be constructed for the target object to ensure the richness of the copywriting through the multi-dimensional copywriting set. At this time, the matching copywriting that matches the user group can be selected from the multi-dimensional copywriting set, so as to ensure that the subsequent processed copywriting is associated with the user group. At the same time, the reference copywriting can be determined in the multi-dimensional copywriting set to ensure that the model can distinguish the preferences of the user group. On this basis, a copywriting generation sample pair can be constructed according to the matching copywriting, and a copywriting preference sample pair can be constructed according to the matching copywriting and the reference copywriting, and the large language model can be optimized based on the copywriting generation sample pair and the copywriting preference sample pair, so that the large language model has the ability to generate copywriting and can learn the preference knowledge of the user group. Finally, the optimized large language model can be used to generate the target copywriting corresponding to the target object. Furthermore, it is ensured that when constructing copywriting for the user group, the user attributes and the generation ability of the large language model can be fully combined, so that the target copywriting better matches the user group, thereby improving the copywriting push effect of downstream services.

[0028] See Figure 2 , Figure 2 FIG. shows a flowchart of a copywriting generation method provided according to an embodiment of the present specification, which specifically includes the following steps.

[0029] Step S202, determine the target object and the user group associated with the target object.

[0030] The copywriting generation method provided in this embodiment can be applied to the short copywriting generation scenario in any service scenario, and is used to push the generated copywriting to the matching user group, thereby improving the push effect. For example, product copywriting push scenarios, game copywriting push scenarios, transaction discount copywriting push scenarios, shopping copywriting push scenarios, article selection copywriting push scenarios, etc. It is used to generate different target copywriting for different user groups, so as to ensure that the generated copywriting better matches the interest preferences of the user group, so as to effectively improve the copywriting push effect and reach more users.

[0031] Taking the application of the copywriting generation method in the product copywriting push scenario as an example, this embodiment describes the copywriting generation method. For the descriptions of other scenarios, reference can be made to the same or corresponding description content in this embodiment, and this embodiment will not elaborate here.

[0032] Specifically, the target object specifically refers to the object for which copywriting needs to be generated in the current scenario, which can be a product to be recommended, a theme to be recommended, an article to be recommended, a game to be recommended, a coupon to be recommended, etc. This embodiment does not make any limitations here. Correspondingly, the user group specifically refers to the group of users who can push the copywriting associated with the target object in the current scenario, and the user group is determined by dividing multiple users included in the current scenario according to different granularities, such as dividing by gender, dividing by occupation, etc. Different user groups have different attributes, so as to ensure that when generating copywriting for this user group subsequently, the copywriting is more in line with the interest preferences of this user group and the user reach rate is improved.

[0033] Based on this, in order to ensure that the generated copywriting is more in line with the user group and has a better effect, the target object and the user group associated with the target object can be determined first, and then a multi-dimensional copywriting set can be constructed for the target object to ensure the richness of the copywriting through the multi-dimensional copywriting set. At this time, the matching copywriting that matches the user group can be selected from the multi-dimensional copywriting set, so as to ensure that the subsequent processed copywriting is associated with the user group. At the same time, the reference copywriting can be determined in the multi-dimensional copywriting set to ensure that the model can distinguish the preferences of the user group. On this basis, a copywriting generation sample pair can be constructed according to the matching copywriting, and a copywriting preference sample pair can be constructed according to the matching copywriting and the reference copywriting, and the large language model can be optimized based on the copywriting generation sample pair and the copywriting preference sample pair, so that the large language model has the ability to generate copywriting and can learn the preference knowledge of the user group. Finally, the optimized large language model can be used to generate the target copywriting corresponding to the target object.

[0034] Furthermore, in order to be able to generate copywriting with different representations for different user groups, thereby improving the push effect on this user group, the user group can be divided into multiple user groups according to the division rules, and then the copywriting generation process can be carried out. In this embodiment, the specific implementation method is as follows: Determine multiple users associated with the target object; divide the multiple users according to the preset division rules to obtain at least two initial user groups, and use the at least two initial user groups as the user groups associated with the target object.

[0035] Specifically, the multiple users specifically refer to the users who are associated with the target object in the current scenario and to whom copywriting needs to be pushed. Correspondingly, the division rules specifically refer to the rules preset in the current scenario for dividing multiple users, which can be set according to different service requirements, such as gender division rules, age division rules, occupation division rules, etc. Specifically, when implementing, the preset division rules and the granularity corresponding to the division rules can be set according to actual needs, and this embodiment does not make any limitations here.

[0036] Based on this, in order to be able to push the copywriting corresponding to the target object to different user groups and ensure that the copywriting reflects personalization, multiple users associated with the target object can be determined first; thereafter, the multiple users can be divided according to a preset division rule to obtain at least two initial user groups according to the division result, and then the at least two initial user groups can be used as the user groups associated with the target object, so as to generate copywriting for each user group respectively for subsequent pushing.

[0037] It should be noted that the multiple users can be all users in the current scenario, or users selected from all users who may be interested in the target object. At the same time, the number of users included in the user group can be set by the division rule. The smaller the rule granularity, the fewer the number of users included in the user group, and it can support a user group containing only one user, thereby further improving the copywriting pushing effect.

[0038] For example, when a shopping platform needs to push copywriting related to mobile phone - A to users, it can first determine multiple users who need to be pushed the copywriting of mobile phone A among the users of the shopping platform, and then divide the multiple users into a male user group and a female user group according to gender. On this basis, subsequent copywriting related to mobile phone A can be constructed for the male user group and the female user group respectively, thereby improving the copywriting pushing effect and ensuring the personalization of the copywriting.

[0039] In summary, by dividing users into multiple user groups according to the set rules, it is possible to combine the attributes corresponding to the user groups to complete the copywriting generation when generating copywriting subsequently, thereby improving the personalization of the copywriting generation.

[0040] Step S204, construct a multi - dimensional copywriting set for the target object, select a matching copywriting that matches the user group from the multi - dimensional copywriting set, and determine a reference copywriting in the multi - dimensional copywriting set.

[0041] Specifically, after determining the target object and its associated user groups above, in order to effectively improve the pushing effect of the copywriting generated subsequently, a multi - dimensional copywriting set can be constructed for the target object first, providing rich copywriting as the basis for target copywriting generation through the multi - dimensional copywriting set. Then, a matching copywriting that matches the user group can be selected from the multi - dimensional copywriting set, and a reference copywriting that is not relevant to the user group can be determined in the multi - dimensional copywriting set, so as to optimize the large - language model by combining the reference copywriting and the matching copywriting, enabling the large - language model to have the ability to generate copywriting and understand the preferences of the current user group at the same time, thereby making the target copywriting generation effect better.

[0042] Among them, the multi-dimensional copywriting set specifically refers to a set composed of copywriting generated in different dimensions. The copywriting included in this set comes from at least two dimensions, which is used to improve the richness of the copywriting in the set, facilitate subsequent matching with user groups, and thus reflect the types of copywriting that user groups are interested in. This enables the large language model to fully understand the attributes of user groups when generating copywriting and improve the accuracy of copywriting generation. Correspondingly, the matching copywriting specifically refers to the copywriting in the multi-dimensional copywriting set that is associated with the user group, that is, the matching copywriting is the copywriting that the user group is interested in. Correspondingly, the reference copywriting specifically refers to the copywriting in the multi-dimensional copywriting set that the user is not interested in.

[0043] Furthermore, when constructing the multi-dimensional copywriting set, in order to ensure that the included copywriting is more abundant, the construction of the copywriting set can be completed by combining the style, style, and theme dimensions. In this embodiment, the specific implementation method is as follows: Determine the preset copywriting style dimension, copywriting theme dimension, and copywriting style dimension; construct style copywriting corresponding to the copywriting style dimension, theme copywriting corresponding to the copywriting theme dimension, and style copywriting corresponding to the copywriting style dimension for the target object; generate a multi-dimensional copywriting set based on the style copywriting, the theme copywriting, and the style copywriting.

[0044] Specifically, the copywriting style dimension specifically refers to the dimension in which copywriting can be generated for user groups according to the copywriting style, which is used to ensure that the generated style copywriting can reflect different styles in terms of language expression, structural layout, emotional color, and creative techniques. Correspondingly, the copywriting theme dimension specifically refers to the dimension in which copywriting is generated for user groups according to the copywriting theme type, which is used to ensure that the generated style copywriting can be generated according to different themes (such as product details, product reviews, product evaluations, etc.). Correspondingly, the copywriting style dimension specifically refers to the dimension in which copywriting is generated for user groups according to the copywriting style, which is used to ensure that the generated style copywriting can be generated according to different styles (such as font color, combination of text and graphics, combination of text and audio, etc.).

[0045] Based on this, when constructing the multi-dimensional copywriting set, in order to improve the richness of the copywriting set, the preset copywriting style dimension, copywriting theme dimension, and copywriting style dimension can be determined first; at this time, style copywriting corresponding to the copywriting style dimension, theme copywriting corresponding to the copywriting theme dimension, and style copywriting corresponding to the copywriting style dimension can be constructed for the target object; finally, the style copywriting, theme copywriting, and style copywriting are integrated to generate a multi-dimensional copywriting set for subsequent use in model optimization. Among them, the copywriting style dimension, copywriting theme dimension, and copywriting style dimension satisfy an orthogonal relationship, so as to ensure that most user preferences for copywriting are covered. Finally, the generated multi-dimensional file set can contain a candidate set of copywriting with different themes, different styles, and different styles.

[0046] In practical applications, for the copywriting style dimension, it is possible to generate copywriting in different styles for different user groups. It can adopt few-shot prompt engineering to generate a large number of copywriting with different styles for use. For the copywriting theme dimension, it is possible to generate copywriting of different theme types for different user groups. For example, for multiple channels (such as product reviews, product details, etc.), the selling points are mined, and the copywriting generation is optimized based on the combined selling points, so as to obtain the theme copywriting corresponding to different types of themes. For the copywriting style dimension, it is possible to generate copywriting in different styles for different user groups. For example, the font color of the copywriting is changed, and the font of many selling points is changed, etc. At the same time, corresponding emojis can be generated according to the semantics and displayed, so as to improve the richness of the generated style copywriting. So that the final multi-dimensional copywriting collection covers multiple dimensions of different user preferences.

[0047] In addition, when generating copywriting for each dimension separately, the open-source dataset and the dataset related to the service project can be used to fine-tune the large language model, so that the model has the basic copywriting generation ability. After that, the fine-tuned model can be used to generate copywriting for each dimension separately, and the copywriting generation for each dimension can be achieved by adding the set prompt words.

[0048] In summary, by generating copywriting for different dimensions separately, the richness of the multi-dimensional copywriting collection can be effectively improved, so as to cover more users, so that when screening copywriting for different user groups subsequently, the selectivity can be improved.

[0049] Furthermore, when selecting matching copywriting and determining the reference copywriting, in order to ensure that the matching copywriting meets the personalized needs of the user group, a matching algorithm can be used to complete the screening of the matching copywriting, and at the same time, the unselected copywriting is used as the reference text for subsequent copywriting generation processing. In this embodiment, the specific implementation method is as follows: Calculate the matching degree between the copywriting in the multi-dimensional copywriting collection and the user group according to the preset matching algorithm; select the copywriting that matches the user group in the multi-dimensional copywriting collection as the matching copywriting based on the matching degree; use the remaining copywriting in the multi-dimensional copywriting collection except the matching copywriting as the reference copywriting.

[0050] Specifically, the preset matching algorithm specifically refers to the offline naive Bandit algorithm, which is used to calculate the matching degree between each copywriting in the multi-dimensional copywriting collection and the user group, and then one or more copywriting can be selected as the matching copywriting for this user group.

[0051] Based on this, in order to ensure that the selected copywriting is more matched to the user group, the matching degree between each piece of copywriting in the multi-dimensional copywriting set and the user group can be calculated according to a preset matching algorithm; then the matching degrees can be sorted, and one or more pieces of copywriting can be selected as the matching copywriting for the user group according to the sorting result; at the same time, the remaining copywriting in the multi-dimensional copywriting set except the matching copywriting can be used as reference copywriting, so as to optimize the large language model by combining the matching copywriting and the reference copywriting in the follow-up.

[0052] In practical applications, when using the offline Naive Bandit algorithm to screen matching copywriting, the algorithm parameters can be set first, such as epsilon and the number of arms. Then several random trials can be carried out on each arm to calculate the average reward of each arm. Further, according to the current average reward and epsilon value, exploitation (select the arm with the largest current reward) or exploration (randomly select an arm) can be chosen. Then the average reward of each arm is updated according to the selected result. And the steps of selection and update evaluation are repeated until the predetermined number of iterations is reached or other stopping conditions are met. The performance of the algorithm can be evaluated using the simulation results, such as calculating the cumulative regret, etc., to obtain the matching degree of each piece of copywriting, and then the copywriting that is more matched to the user group can be screened out as the matching copywriting for subsequent processing.

[0053] Continuing with the above example, after obtaining the male user group and the female user group, for mobile phone A, style copywriting in the dimension of matching copywriting style can be generated first, such as using the large language model LLM to generate style copywriting according to the concise and clear style, vivid and interesting style, emotional resonance style, etc. At this time, n pieces of style copywriting are obtained. Theme copywriting in the dimension of matching copywriting theme can also be generated, such as using the large language model LLM to generate theme copywriting according to the mobile phone details information and mobile phone review information, etc. At this time, m pieces of theme copywriting are obtained. Style copywriting in the dimension of matching copywriting style can also be generated, such as using the large language model LLM to generate style copywriting according to the image combined with text style, video combined with text style, audio combined with text style, etc. At this time, q pieces of style copywriting are obtained. Then by integrating n pieces of style copywriting, m pieces of theme copywriting, and q pieces of style copywriting, the multi-dimensional copywriting set corresponding to mobile phone A can be obtained, and this set contains x candidate copywriting.

[0054] Further, after obtaining the multi-dimensional copywriting set containing x candidate copywritings, the offline Naive Bandit algorithm can be used to select preferred copywritings for male and female user groups respectively. According to the calculation results, determine the candidate copywritings x1, x2, x3, and x4 in the multi-dimensional copywriting set preferred by the male user group, and at the same time determine that the male user group is not interested in the candidate copywriting x5. The female user group prefers the candidate copywritings x1, x2, x3, and x5 in the multi-dimensional copywriting set, and at the same time determines that the female user group is not interested in the candidate copywriting x4. Subsequently, the optimization of the large language model LLM can be completed by combining the above copywritings, and then different copywritings can be constructed for different user groups for pushing.

[0055] In summary, by combining the offline Naive Bandit algorithm to separately select matching copywritings for different user groups, the matching accuracy can be effectively improved, and then the accuracy of the subsequent generated target copywriting can be guaranteed.

[0056] Step S206, constructing a copywriting generation sample pair according to the matching copywriting, and constructing a copywriting preference sample pair according to the matching copywriting and the reference copywriting.

[0057] Specifically, after determining the matching copywriting for the user group and the irrelevant reference copywriting above, in order to achieve the purpose of optimizing the large language model, a copywriting generation sample pair can be constructed according to the matching copywriting for optimizing the large language model, so that the model has the ability to generate copywritings related to the theme, style, and style of the matching copywriting. At the same time, a copywriting preference sample pair can be constructed according to the matching copywriting and the reference copywriting. After optimizing the large language model, the model has the ability to generate copywritings related to the preferences of the user group, thereby improving the generation effect of the target copywriting.

[0058] Among them, the copywriting generation sample pair specifically refers to a sample pair formed by using the matching copywriting as a label and constructing samples. Its function is to enable the model to have the ability to generate copywritings related to the theme, style, and style of the matching copywriting. Correspondingly, the copywriting preference sample pair specifically refers to a sample pair formed by using the matching copywriting and the reference copywriting as labels and constructing samples. Its function is to enable the model to have the ability to generate copywritings related to the preferences of the user group.

[0059] Further, the construction of the copywriting generation sample pair is specifically implemented as follows: Extract the first copywriting element information from the matching copywriting, and construct copywriting generation information based on the first copywriting element information; use the matching copywriting as the copywriting generation label, and the copywriting generation information as the copywriting generation sample; construct a copywriting generation sample pair based on the copywriting generation label and the copywriting generation sample.

[0060] Specifically, the first copywriting element information specifically refers to the element information that matches the associated styles, themes, and styles contained in the copywriting, and is used to construct the model input, that is, the copywriting generation information.

[0061] Based on this, in order to optimize the model through samples so that the model output contains copywriting with styles, styles, and themes that match the user group, the first copywriting element information can be extracted from the matching copywriting, and the copywriting generation information can be constructed based on the first copywriting element information; thereafter, the matching copywriting can be used as the copywriting generation label, and the copywriting generation information can be used as the copywriting generation sample; furthermore, a copywriting generation sample pair can be constructed based on the copywriting generation label and the copywriting generation sample.

[0062] In summary, by using the matching copywriting as a label and constructing a copywriting generation sample, the construction of the copywriting generation sample pair can be completed, enabling the model to have the ability to generate copywriting that matches the current scenario and conforms to the user group during the model optimization stage.

[0063] Furthermore, the construction of the copywriting preference sample pair is specifically implemented as follows: Extract the second copywriting element information from the matching copywriting and the third copywriting element information from the reference copywriting; construct the copywriting preference information according to the second copywriting element information and the third copywriting element information; use the matching copywriting and the reference copywriting as the copywriting preference labels, and the copywriting preference information as the copywriting preference sample; construct a copywriting preference sample pair based on the copywriting preference labels and the copywriting preference sample.

[0064] Specifically, for the description of the second copywriting element information and the third copywriting element information, reference can be made to the description of the first copywriting element information in the above embodiments, and this embodiment does not make any limitations here. Correspondingly, the copywriting preference information specifically refers to the description information that reflects the samples, styles, and themes that the user group is interested in, as well as the samples, styles, and themes that the user group is not interested in.

[0065] Based on this, in order to enable the model to have the ability to predict copywriting preferences, the second copywriting element information can be extracted from the matching copywriting and the third copywriting element information can be extracted from the reference copywriting; at this time, the copywriting preference information can be constructed according to the second copywriting element information and the third copywriting element information; furthermore, by using the matching copywriting and the reference copywriting as the copywriting preference labels, and the copywriting preference information as the copywriting preference sample, the purpose of constructing a copywriting preference sample pair by combining the copywriting preference labels and the copywriting preference sample can be achieved, so as to be used for subsequent model optimization.

[0066] Continuing with the above example, after obtaining the candidate copywriting that male and female user groups prefer respectively and the candidate copywriting that they do not prefer, further, for the two user groups, a candidate copywriting can be randomly selected from their matching candidate copywriting as the copywriting generation label. At the same time, the model input matching the candidate copywriting is constructed. Combining the copywriting generation label and the model input, a copywriting generation sample pair can be constructed; it is also possible to randomly select a candidate copywriting from their matching candidate copywriting, and randomly select a candidate copywriting from the non-matching candidate copywriting, combine the two selected candidate copywriting to generate a copywriting preference label, and at the same time construct the model input. Combining the copywriting preference label and the model input, a copywriting generation sample pair can be constructed.

[0067] For example, select the candidate copywriting x1 as the label, and extract the style information a1, theme information b1, and style information c1 from the candidate copywriting x1; thereafter, the model input y11 = {Please generate a copywriting about mobile phone A for male users, and include the style a1, theme b1, and style c1} can be constructed for the candidate copywriting x1 in combination with the above information. Furthermore, combining the candidate copywriting x1 as the label and the above model input y1, a copywriting generation sample pair is generated.

[0068] At the same time, the candidate copywriting x1 and the candidate copywriting x5 that male user groups are not interested in can also be selected, and a preference label is constructed by combining the candidate copywriting x1 and x5. Then, the style information a5, theme information b5, and style information c5 are extracted from the candidate copywriting x5; on this basis, the model input y15 = {Please generate a copywriting about mobile phone A for male users, and include the style a1, theme b1, and style c1, without the style a5, theme b5, and style c5} is constructed for the candidate copywriting x1 and x5 in combination with the above information. Furthermore, combining the preference label and the above model input y15, a copywriting preference sample pair is generated.

[0069] And so on, for the male user group, a set of male copywriting generation sample pairs and a set of male copywriting preference sample pairs can be obtained. For the female user group, a set of female copywriting generation sample pairs and a set of female copywriting preference sample pairs can be obtained, which is convenient for subsequent fine-tuning of the large language model LLM respectively, so that the model can output different copywriting for different user groups and push them to the corresponding users.

[0070] In summary, by constructing copywriting generation sample pairs and copywriting preference sample pairs, the model can learn the copywriting generation ability regarding user group preferences during the optimization stage, thereby improving the accuracy of the finally generated target copywriting.

[0071] Step S208, optimize the large language model based on the copywriting generation sample pairs and the copywriting preference sample pairs, and use the optimized large language model to generate the target copywriting corresponding to the target object.

[0072] Specifically, after obtaining the copywriting generation sample pairs and copywriting preference sample pairs as described above, the large language model can be fine-tuned using the copywriting generation sample pairs and copywriting preference sample pairs. After the fine-tuned large language model matches the user group, the optimized large language model can be used to generate the target copywriting corresponding to the target object, and it is ensured that the finally generated target copywriting matches the user group, thereby improving the copywriting push effect. Among them, the target copywriting specifically refers to the copywriting that matches the user group.

[0073] Furthermore, in order to make the model output more accurate target copywriting, the large language model can be fine-tuned in combination with the sample pairs and then used in the current scenario. In this embodiment, the specific implementation method is as follows: Optimize the large language model according to the copywriting generation sample pairs and the copywriting preference sample pairs until a target large language model that meets the optimization stop condition is obtained; input the object description information corresponding to the target object into the target large language model for processing to obtain the target copywriting that the target object matches the user group.

[0074] Specifically, the optimization stop condition specifically refers to the condition for stopping the optimization of the large language model, which can be a loss value comparison condition, a validation set verification condition, or an iteration number condition. Specifically in implementation, it can be set according to actual needs, and this embodiment does not make any limitations here. Correspondingly, the object description information specifically refers to the model input information associated with the target object, and this object description information contains the attribute characteristics of the user group, thereby improving the effect of the large language model generating copywriting for this user group.

[0075] Based on this, in the model optimization stage, the samples included in the copywriting generation sample pairs and the copywriting preference sample pairs can be used as model inputs respectively. After being processed by the model, the prediction results corresponding to each sample will be obtained. Then, the loss value is calculated using the labels in the sample pairs and the prediction results, and the model is fine-tuned according to the calculated loss value. Iterate in this way until the model meets the optimization stop condition after a certain fine-tuning process, and then it can be used as the target large language model. Further, by constructing object description information that matches the user group and inputting it into the target large language model for processing, the target copywriting that the target object output by the model matches the user group can be obtained.

[0076] Continuing with the above example, when generating copywriting for the female user group, the large language model LLM can be fine-tuned using the obtained female copywriting generation sample pair set and female copywriting preference sample pair set. When the fine-tuned model meets the set conditions, the model input can be constructed by combining the attribute information of the female user group {Generate a copywriting about mobile phone A for female users, and it contains a 女 Style a 女 Theme and a 女Style}, after being processed by the fine-tuned model, the generated copywriting 1 is {As a brand-new mobile phone of brand A, the camera configuration of the A mobile phone has been further improved, supporting a brand-new AI photo retouching function, which can save you more trouble in photo retouching...}. After that, this copywriting 1 can be sent to the terminal devices of female users included in the female user group, thus completing the copywriting push.

[0077] When generating copywriting for the male user group, the obtained male copywriting generation sample pair set and male copywriting preference sample pair set can be used to fine-tune the large language model LLM. When the fine-tuned model meets the set conditions, it can combine the attribute information of the male user group to construct the model input {Please generate a copywriting about the A mobile phone for male users and include a 男 Style a 男 Theme and a 男 Style}, after being processed by the fine-tuned model, the generated copywriting 2 is {As a brand-new mobile phone of brand A, the A mobile phone is equipped with a newly developed processor, supports users to simply overclock, and at the same time the screen resolution reaches...}. After that, this copywriting 2 can be sent to the terminal devices of male users included in the male user group, thus completing the copywriting push.

[0078] In summary, by selecting sample pairs that match the user group to optimize the large language model, enabling it to have the ability to generate copywriting that matches the user group's preferences, the target copywriting generated based on this can better match the user group, thereby effectively improving the push effect of short copywriting.

[0079] In addition, for newly added copywriting, the above process can be repeated to further improve the copywriting generation accuracy of the large language model. In this embodiment, the specific implementation method is as follows: When the newly added copywriting is obtained, determine the target user group that the newly added copywriting matches; use the newly added copywriting as the matching copywriting that matches the target user group, and perform the steps of constructing a copywriting generation sample pair according to the matching copywriting, and constructing a copywriting preference sample pair according to the matching copywriting and the reference copywriting.

[0080] Specifically, the newly added copywriting specifically refers to the copywriting newly added in the current scenario, which can be set manually or generated through the above processing process. Correspondingly, the target user group specifically refers to the user group that matches the newly added copywriting.

[0081] Based on this, when new copywriting is obtained, in order to enable the large language model to quickly learn new knowledge, the target user group that the new copywriting matches can be determined first; at this time, the new copywriting can be used as the matching copywriting for the target user group, and the steps of constructing a copywriting generation sample pair according to the matching copywriting and constructing a copywriting preference sample pair according to the matching copywriting and the reference copywriting can be returned for execution, so as to achieve the purpose of sustainable automatic optimization to meet the actual application scenario.

[0082] For the copywriting generation method provided in this embodiment, in order to ensure that the generated copywriting is more matched to the user group and has a better effect, the target object and the user group associated with the target object can be determined first, and then a multi-dimensional copywriting set can be constructed for the target object first to ensure the richness of the copywriting through the multi-dimensional copywriting set. At this time, the matching copywriting that matches the user group can be selected from the multi-dimensional copywriting set, so as to ensure that the subsequent processed copywriting is associated with the user group. At the same time, the reference copywriting can be determined from the multi-dimensional copywriting set to ensure that the model can distinguish the preferences of the user group. On this basis, a copywriting generation sample pair can be constructed according to the matching copywriting, and a copywriting preference sample pair can be constructed according to the matching copywriting and the reference copywriting, and the large language model can be optimized based on the copywriting generation sample pair and the copywriting preference sample pair, so that the large language model has the ability to generate copywriting and can learn the preference knowledge of the user group. Finally, the optimized large language model can be used to generate the target copywriting corresponding to the target object. Furthermore, it is ensured that when constructing copywriting for the user group, the user attributes and the generation ability of the large language model can be fully combined, so that the target copywriting is more matched to the user group, thereby improving the copywriting push effect of downstream services.

[0083] See Figure 3 , Figure 3 shows a flowchart of another copywriting generation method provided according to an embodiment of this specification, which specifically includes the following steps.

[0084] Step S302, determine the target commodity and the user group associated with the target commodity.

[0085] Step S304, construct a multi-dimensional push copywriting set for the target commodity, select the matching copywriting that matches the user group from the multi-dimensional push copywriting set, and determine the reference copywriting from the multi-dimensional push copywriting set.

[0086] Step S306, construct a copywriting generation sample pair according to the matching copywriting, and construct a copywriting preference sample pair according to the matching copywriting and the reference copywriting.

[0087] Step S308, optimize the large language model based on the copywriting generation sample pair and the copywriting preference sample pair, and use the optimized large language model to generate the target push copywriting corresponding to the target commodity.

[0088] Step S310: Send the target push copywriting to the terminal devices of the users included in the user group.

[0089] Another copywriting generation method provided in this embodiment is applied to the push scenario of commodity copywriting. For the same or corresponding description content in the above embodiments, reference can be made to the above embodiments, and this embodiment will not be elaborated here too much. Among them, the target commodity specifically refers to the commodity related to the message push to users in the online shopping platform, which can be mobile phones, computers, cosmetics, toys, cars, daily necessities, etc. Correspondingly, the terminal device specifically refers to the device held by the users in the user group, which can be used to receive the target push copywriting, such as intelligent devices such as computers, mobile phones, and tablets.

[0090] Based on this, in order to be able to push different copywritings for different user groups for the target commodity, thereby improving the reach rate, the target commodity and the user group associated with the target commodity can be determined first; at this time, a multi-dimensional push copywriting set can be constructed for the target commodity first, and a matching copywriting that matches the user group can be selected from the multi-dimensional push copywriting set, and a reference copywriting can be determined from the multi-dimensional push copywriting set; furthermore, a copywriting generation sample pair can be constructed by using the matching copywriting, and a copywriting preference sample pair can be constructed by using the matching copywriting and the reference copywriting; after optimizing the large language model in this way, the large language model can have the ability to generate copywriting that meets the preferences for the user group, so the target push copywriting corresponding to the target commodity can be generated by using the optimized large language model; finally, the target push copywriting is sent to the terminal devices of the users included in the user group, and the push effect of the copywriting associated with the target commodity can be effectively improved.

[0091] In summary, through the copywriting generation method provided in this embodiment, personalized push copywriting of associated commodities can be generated for different user groups, thereby effectively improving the push effect of the copywriting and ensuring that the copywriting can reach more users.

[0092] See Figure 4 , Figure 4 shows a flowchart of a multi-modal information generation method provided according to an embodiment of this specification, which specifically includes the following steps.

[0093] Step S402: Determine the target object and the user group associated with the target object.

[0094] Step S404: Construct a multi-dimensional multi-modal information set for the target object, select matching multi-modal information that matches the user group from the multi-dimensional multi-modal information set, and determine reference multi-modal information from the multi-dimensional multi-modal information set.

[0095] Step S406: Construct multimodal information generation sample pairs according to the matched multimodal information, and construct multimodal information preference sample pairs according to the matched multimodal information and the reference multimodal information.

[0096] Step S408: Optimize the large language model based on the multimodal information generation sample pairs and the multimodal information preference sample pairs, and use the optimized large language model to generate the target multimodal information corresponding to the target object.

[0097] The multimodal information generation method provided in this embodiment can be applied to generate multimodal information of the associated user group for any object. For the same or corresponding description content in the above embodiments, reference can be made to the above embodiments, and this embodiment will not be elaborated here. Among them, the multi-dimensional multimodal information set can refer to the description of the multi-dimensional copy set, the matched multimodal information can refer to the description of the matched copy; the reference multimodal information can refer to the description of the reference copy; the multimodal information generation sample pair can refer to the description of the copy generation sample pair; the multimodal information preference sample pair can refer to the description of the copy preference sample pair; the target multimodal information can refer to the description of the target copy.

[0098] Based on this, in order to ensure that the generated target multimodal information better matches the user group, the target object and the user group associated with the target object can be determined first, and then a multi-dimensional multimodal information set can be constructed for the target object to ensure the richness of multimodal information through the multi-dimensional multimodal information set. At this time, the matched multimodal information that matches the user group can be selected from the multi-dimensional multimodal information set, so as to ensure that the subsequent processed multimodal information is associated with the user group. At the same time, the reference multimodal information can be determined in the multi-dimensional multimodal information set to ensure that the model can distinguish the preferences of the user group. On this basis, multimodal information generation sample pairs can be constructed according to the matched multimodal information, and multimodal information preference sample pairs can be constructed according to the matched multimodal information and the reference multimodal information, and the large language model can be optimized based on the multimodal information generation sample pairs and the multimodal information preference sample pairs, so that the large language model has the ability to generate multimodal information and can learn the preference knowledge of the user group. Finally, the optimized large language model can be used to generate the target multimodal information corresponding to the target object.

[0099] For example, when generating graphic and text description information for computer popularization knowledge videos, the user group watching the video can be determined. Then, various types of graphic and text description information can be generated for the computer popularization knowledge videos from multiple dimensions and combined into a multi-dimensional graphic and text description information set. Further, the graphic and text description information that matches the user group can be selected from the multi-dimensional graphic and text description information set, and the remaining other graphic and text description information can be used as reference graphic and text description information. Furthermore, an information generation sample pair can be constructed by combining the graphic and text description information that matches the user group, and an information preference sample pair can be constructed by combining the reference graphic and text description information. After that, the large language model can be fine-tuned using the information generation sample pair and the information preference sample pair. After the model is fine-tuned, the description text of the computer popularization knowledge video can be input into the large language model for processing, and the target graphic and text explanation information about the computer popularization knowledge video output by the large language model can be obtained according to the processing result. Finally, the target graphic and text explanation information can be sent to the above-mentioned users, thereby achieving the purpose of popularizing computer knowledge.

[0100] In summary, through the multi-modal information generation method provided in this embodiment, it is possible to generate personalized multi-modal information of associated objects for different user groups, thereby effectively ensuring the accuracy of information generation and meeting the browsing needs of user groups.

[0101] The following combines the attached Figure 5 , taking the application of the copywriting generation method provided in this specification in the commodity message push scenario as an example, to further illustrate the copywriting generation method. Among them, Figure 5 Fig. shows the processing procedure flowchart of a copywriting generation method provided in an embodiment of this specification, which specifically includes the following steps.

[0102] Step S502, determine the target object and the user group associated with the target object, and determine the preset copywriting style dimension, copywriting theme dimension, and copywriting style dimension.

[0103] Step S504, construct a style copywriting corresponding to the copywriting style dimension, a theme copywriting corresponding to the copywriting theme dimension, and a style copywriting corresponding to the copywriting style dimension for the target object.

[0104] Step S506, generate a multi-dimensional copywriting set based on the style copywriting, theme copywriting, and style copywriting.

[0105] Step S508, calculate the matching degree between the copywriting in the multi-dimensional copywriting set and the user group according to the preset matching algorithm.

[0106] Step S510, select the copywriting that matches the user group in the multi-dimensional copywriting set as the matching copywriting based on the matching degree.

[0107] Step S512: Use the remaining copywriting in the multi-dimensional copywriting set except the matching copywriting as reference copywriting.

[0108] Step S514: Extract the first copywriting element information from the matching copywriting and construct copywriting generation information based on the first copywriting element information.

[0109] Step S516: Use the matching copywriting as the copywriting generation label and the copywriting generation information as the copywriting generation sample.

[0110] Step S518: Construct a copywriting generation sample pair based on the copywriting generation label and the copywriting generation sample, and extract the second copywriting element information from the reference copywriting.

[0111] Step S520: Construct copywriting preference information according to the first copywriting element information and the second copywriting element information.

[0112] Step S522: Use the matching copywriting and the reference copywriting as the copywriting preference label, and the copywriting preference information as the copywriting preference sample.

[0113] Step S524: Construct a copywriting preference sample pair based on the copywriting preference label and the copywriting preference sample.

[0114] Step S526: Optimize the large language model according to the copywriting generation sample pair and the copywriting preference sample pair until the target large language model that meets the optimization stop condition is obtained.

[0115] Step S528: Input the object description information corresponding to the target object into the target large language model for processing to obtain the target copywriting that matches the user group of the target object.

[0116] In summary, in order to ensure that the generated copywriting is more suitable for the user group and has better effects, the target object and the user group associated with the target object can be determined first. Then, a multi-dimensional copywriting set can be constructed for the target object to ensure the richness of the copywriting through the multi-dimensional copywriting set. At this time, the matching copywriting that matches the user group can be selected from the multi-dimensional copywriting set, so as to ensure that the subsequent processed copywriting is associated with the user group. At the same time, the reference copywriting can be determined in the multi-dimensional copywriting set to ensure that the model can distinguish the preferences of the user group. On this basis, a copywriting generation sample pair can be constructed according to the matching copywriting, and a copywriting preference sample pair can be constructed according to the matching copywriting and the reference copywriting. The large language model can be optimized based on the copywriting generation sample pair and the copywriting preference sample pair, so that the large language model has the ability to generate copywriting and can learn the preference knowledge of the user group. Finally, the optimized large language model can be used to generate the target copywriting corresponding to the target object. Furthermore, it is ensured that when constructing copywriting for the user group, the user attributes and the generation ability of the large language model can be fully combined, so that the target copywriting is more suitable for the user group, thereby improving the copywriting push effect of downstream services.

[0117] Corresponding to the above method embodiment, this specification also provides an embodiment of a copywriting generation device. Figure 6 The structure diagram of a copywriting generation device provided by an embodiment of this specification is shown. As Figure 6 shown, the device includes: A determination module 602, configured to determine a target object and the user group associated with the target object; A selection module 604, configured to construct a multi-dimensional copywriting set for the target object, select the matching copywriting that matches the user group from the multi-dimensional copywriting set, and determine the reference copywriting in the multi-dimensional copywriting set; A construction module 606, configured to construct a copywriting generation sample pair according to the matching copywriting, and construct a copywriting preference sample pair according to the matching copywriting and the reference copywriting; A generation module 608, configured to optimize the large language model based on the copywriting generation sample pair and the copywriting preference sample pair, and use the optimized large language model to generate the target copywriting corresponding to the target object.

[0118] In an optional embodiment, the selection module 604 is further configured to: Determine the preset copywriting style dimension, copywriting theme dimension, and copywriting style dimension; construct the style copywriting corresponding to the copywriting style dimension, the theme copywriting corresponding to the copywriting theme dimension, and the style copywriting corresponding to the copywriting style dimension for the target object; generate a multi-dimensional copywriting set based on the style copywriting, the theme copywriting, and the style copywriting.

[0119] In an alternative embodiment, the selection module 604 is further configured to: Calculate the matching degree between the copywriting in the multi-dimensional copywriting set and the user group according to a preset matching algorithm; select the copywriting that matches the user group from the multi-dimensional copywriting set based on the matching degree as the matching copywriting; and use the remaining copywriting in the multi-dimensional copywriting set except the matching copywriting as the reference copywriting.

[0120] In an alternative embodiment, the construction module 606 is further configured to: Extract the first copywriting element information from the matching copywriting, and construct copywriting generation information based on the first copywriting element information; use the matching copywriting as the copywriting generation label, and the copywriting generation information as the copywriting generation sample; and construct a copywriting generation sample pair based on the copywriting generation label and the copywriting generation sample.

[0121] In an alternative embodiment, the construction module 606 is further configured to: Extract the second copywriting element information from the matching copywriting and the third copywriting element information from the reference copywriting; construct copywriting preference information according to the second copywriting element information and the third copywriting element information; use the matching copywriting and the reference copywriting as the copywriting preference label, and the copywriting preference information as the copywriting preference sample; and construct a copywriting preference sample pair based on the copywriting preference label and the copywriting preference sample.

[0122] In an alternative embodiment, the generation module 608 is further configured to: Optimize the large language model according to the copywriting generation sample pair and the copywriting preference sample pair until a target large language model that meets the optimization stop condition is obtained; input the object description information corresponding to the target object into the target large language model for processing, and obtain the target copywriting that the target object matches the user group.

[0123] In an alternative embodiment, the determination of the user group includes: Determine multiple users associated with the target object; divide the multiple users according to a preset division rule to obtain at least two initial user groups, and use the at least two initial user groups as the user groups associated with the target object.

[0124] In an alternative embodiment, the device further includes: The group determination module is configured to determine a target user group that matches the newly added copywriting when the newly added copywriting is obtained; use the newly added copywriting as the matching copywriting for the target user group, and perform steps of constructing a copywriting generation sample pair according to the matching copywriting and constructing a copywriting preference sample pair according to the matching copywriting and the reference copywriting.

[0125] In order to ensure that the generated copywriting is more matched to the user group and has a better effect, the copywriting generation device provided in this embodiment can first determine the target object and the user group associated with the target object, and then construct a multi-dimensional copywriting set for the target object to ensure the richness of the copywriting through the multi-dimensional copywriting set. At this time, the matching copywriting that matches the user group can be selected from the multi-dimensional copywriting set, so as to ensure that the subsequent processed copywriting is associated with the user group. At the same time, the reference copywriting can be determined from the multi-dimensional copywriting set to ensure that the model can distinguish the preferences of the user group. On this basis, a copywriting generation sample pair can be constructed according to the matching copywriting, and a copywriting preference sample pair can be constructed according to the matching copywriting and the reference copywriting, and the large language model can be optimized based on the copywriting generation sample pair and the copywriting preference sample pair, so that the large language model has the ability to generate copywriting and can learn the preference knowledge of the user group. Finally, the optimized large language model can be used to generate the target copywriting corresponding to the target object. Furthermore, it is ensured that when constructing copywriting for the user group, the user attributes and the generation ability of the large language model can be fully combined, so that the target copywriting is more matched to the user group, thereby improving the copywriting push effect of downstream services.

[0126] The above is a schematic solution of a copywriting generation device in this embodiment. It should be noted that the technical solution of this copywriting generation device and the technical solution of the above copywriting generation method belong to the same concept. For the details not described in the technical solution of the copywriting generation device, reference can be made to the description of the technical solution of the above copywriting generation method.

[0127] Corresponding to the above method embodiment, this specification also provides another embodiment of a copywriting generation device. Figure 7 Fig. shows a structural schematic diagram of another copywriting generation device provided in an embodiment of this specification. As Figure 7 shown, the device includes: A commodity determination module 702, configured to determine a target commodity and a user group associated with the target commodity; A copywriting selection module 704, configured to construct a multi-dimensional push copywriting set for the target commodity, select a matching copywriting that matches the user group from the multi-dimensional push copywriting set, and determine a reference copywriting from the multi-dimensional push copywriting set; A sample building module 706 is configured to build a copywriting generation sample pair according to the matching copywriting, and build a copywriting preference sample pair according to the matching copywriting and the reference copywriting; A copywriting generation module 708 is configured to optimize a large language model based on the copywriting generation sample pair and the copywriting preference sample pair, and use the optimized large language model to generate a target push copywriting corresponding to the target commodity; A push copywriting module 710 is configured to send the target push copywriting to the terminal devices of the users included in the user group.

[0128] The above is a schematic solution of another copywriting generation device of this embodiment. It should be noted that the technical solution of this another copywriting generation device and the technical solution of the above another copywriting generation method belong to the same concept. For the details not described in detail in the technical solution of this another copywriting generation device, reference can be made to the description of the technical solution of the above another copywriting generation method.

[0129] Corresponding to the above method embodiment, this specification also provides an embodiment of a multimodal information generation device. Figure 8 The structural schematic diagram of a multimodal information generation device provided by an embodiment of this specification is shown. As Figure 8 shown, the device includes: A group determination module 802 is configured to determine a target object and a user group associated with the target object; A selection information module 804 is configured to build a multi-dimensional multimodal information set for the target object, select matching multimodal information that matches the user group from the multi-dimensional multimodal information set, and determine reference multimodal information in the multi-dimensional multimodal information set; A sample building module 806 is configured to build a multimodal information generation sample pair according to the matching multimodal information, and build a multimodal information preference sample pair according to the matching multimodal information and the reference multimodal information; An information generation module 808 is configured to optimize a large language model based on the multimodal information generation sample pair and the multimodal information preference sample pair, and use the optimized large language model to generate target multimodal information corresponding to the target object.

[0130] The above is a schematic solution of a multimodal information generation device of this embodiment. It should be noted that the technical solution of this multimodal information generation device and the technical solution of the above multimodal information generation method belong to the same concept. For the details not described in detail in the technical solution of this multimodal information generation device, reference can be made to the description of the technical solution of the above multimodal information generation method.

[0131] Figure 9FIG. 0 shows a structural block diagram of a computing device 900 provided according to an embodiment of the present specification. The components of the computing device 900 include, but are not limited to, a memory 910 and a processor 920. The processor 920 is connected to the memory 910 via a bus 930, and a database 950 is used to store data.

[0132] The computing device 900 further includes an access device 940, which enables the computing device 900 to communicate via one or more networks 960. Examples of such networks include the Public Switched Telephone Network (PSTN), Local Area Network (LAN), Wide Area Network (WAN), Personal Area Network (PAN), or a combination of communication networks such as the Internet. The access device 940 may include one or more of any type of wired or wireless network interfaces (e.g., a network interface card (NIC)), such as an IEEE 802.11 Wireless Local Area Network (WLAN) wireless interface, a Worldwide Interoperability for Microwave Access (Wi-MAX) interface, an Ethernet interface, a Universal Serial Bus (USB) interface, a cellular network interface, a Bluetooth interface, a Near Field Communication (NFC).

[0133] In an embodiment of the present specification, the above components of the computing device 900 and Figure 9 other components not shown may also be connected to each other, for example, via a bus. It should be understood that Figure 9 the shown structural block diagram of the computing device is for illustrative purposes only and is not a limitation on the scope of the present specification. Those skilled in the art may add or replace other components as needed.

[0134] The computing device 900 can be any type of stationary or mobile computing device, including mobile computers or mobile computing devices (e.g., tablet computers, personal digital assistants, laptop computers, notebook computers, netbooks, etc.), mobile phones (e.g., smart phones), wearable computing devices (e.g., smart watches, smart glasses, etc.) or other types of mobile devices, or stationary computing devices such as desktop computers or personal computers (PCs). The computing device 900 can also be a mobile or stationary server.

[0135] Among them, the processor 920 is used to execute the following computer-executable instructions, and when the computer-executable instructions are executed by the processor, the steps of the above-mentioned copywriting generation method or multi-modal information generation method are implemented.

[0136] The above is a schematic solution of a computing device in this embodiment. It should be noted that the technical solution of this computing device and the technical solutions of the above copywriting generation method or multi-modal information generation method belong to the same concept. For the detailed content not described in the technical solution of the computing device, reference can be made to the description of the technical solutions of the above copywriting generation method or multi-modal information generation method.

[0137] An embodiment of this specification also provides a computer-readable storage medium, which stores computer-executable instructions, and when the computer-executable instructions are executed by a processor, the steps of the above copywriting generation method or multi-modal information generation method are implemented.

[0138] The above is a schematic solution of a computer-readable storage medium in this embodiment. It should be noted that the technical solution of this storage medium and the technical solutions of the above copywriting generation method or multi-modal information generation method belong to the same concept. For the detailed content not described in the technical solution of the storage medium, reference can be made to the description of the technical solutions of the above copywriting generation method or multi-modal information generation method.

[0139] An embodiment of this specification also provides a computer program, wherein when the computer program is executed on a computer, the computer is made to execute the steps of the above copywriting generation method or multi-modal information generation method.

[0140] The above is a schematic solution of a computer program in this embodiment. It should be noted that the technical solution of this computer program and the technical solutions of the above copywriting generation method or multi-modal information generation method belong to the same concept. For the detailed content not described in the technical solution of the computer program, reference can be made to the description of the technical solutions of the above copywriting generation method or multi-modal information generation method.

[0141] An embodiment of this specification also provides a computer program product, including a computer program or instruction, which, when executed by a processor, implements the steps of the above-mentioned copywriting generation method or multimodal information generation method.

[0142] The above is a schematic solution of a computer program product of this embodiment. It should be noted that the technical solution of this computer program product and the technical solutions of the above-mentioned copywriting generation method or multimodal information generation method belong to the same concept. For the details not described in detail in the technical solution of the computer program product, reference can be made to the descriptions of the technical solutions of the above-mentioned copywriting generation method or multimodal information generation method.

[0143] The above describes specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than in the embodiments and still achieve the desired result. Additionally, the processes depicted in the figures do not necessarily require the specific order or sequential order shown to achieve the desired result. In certain embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0144] The computer instructions include computer program code, which may be in source code form, object code form, executable file or some intermediate form, etc. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording medium, USB flash drive, mobile hard disk, magnetic disk, optical disc, computer memory, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), electrical carrier signal, telecommunication signal, and software distribution medium, etc. It should be noted that the content included in the computer-readable medium can be appropriately increased or decreased according to the requirements of patent practice. For example, in some regions, according to patent practice, the computer-readable medium does not include electrical carrier signals and telecommunication signals.

[0145] It should be noted that for the foregoing method embodiments, for the sake of simplicity of description, they are all expressed as a series of action combinations. However, those skilled in the art should know that the embodiments of this specification are not limited by the described action sequence, because according to the embodiments of this specification, certain steps can be performed in other sequences or simultaneously. Secondly, those skilled in the art should also know that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily essential to the embodiments of this specification.

[0146] In the above embodiments, the descriptions of each embodiment have their own focuses. For the parts not detailed in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0147] The preferred embodiments of the present specification disclosed above are only used to help explain the present specification. The alternative embodiments do not describe all the details in detail, nor do they limit the invention to the specific embodiments described. Obviously, according to the content of the embodiments of the present specification, many modifications and variations can be made. These embodiments are selected and specifically described in this specification in order to better explain the principles and practical applications of the embodiments of the present specification, so that those skilled in the art can well understand and utilize this specification. This specification is only limited by the claims and their full scope and equivalents.

Claims

1. A copywriting generation method, comprising: Determining a target object and a user group associated with the target object; Constructing a multi-dimensional copy set for the target object, selecting a matching copy that matches the user group from the multi-dimensional copy set, and determining a reference copy from the multi-dimensional copy set; Constructing a copy generation sample pair according to the matching copy, and constructing a copy preference sample pair according to the matching copy and the reference copy; The large language model is optimized based on the copy generation sample pairs and the copy preference sample pairs, and the target copy corresponding to the target object is generated using the optimized large language model.

2. The method for generating a copywriting according to claim 1, wherein the step of constructing a multi-dimensional copywriting set for the target object comprises: Determine the preset copywriting style dimension, copywriting theme dimension, and copywriting style dimension; Constructing, for the target object, a style copy corresponding to the copy style dimension, a theme copy corresponding to the copy theme dimension, and a style copy corresponding to the copy style dimension; A multi-dimensional copy set is generated based on the style copy, the theme copy and the pattern copy.

3. The method for generating a copy according to claim 1, wherein selecting a matching copy that matches the user group from the multi-dimensional copy set and determining a reference copy from the multi-dimensional copy set comprises: Calculating the matching degree between the copy in the multi-dimensional copy set and the user group according to a preset matching algorithm; Selecting, from the multi-dimensional copy set, a copy that matches the user group as a matching copy based on the matching degree; The remaining documents in the multi-dimensional document set except the matching document are used as reference documents.

4. The method for generating a copywriting according to claim 1, wherein the step of constructing a copywriting sample pair based on the matching copywriting comprises: Extracting first text element information from the matching text, and constructing text generation information based on the first text element information; Using the matching text as a text generation label and the text generation information as a text generation sample; A copy generation sample pair is constructed based on the copy generation label and the copy generation sample.

5. The copywriting generation method according to claim 1, wherein the step of constructing a copywriting preference sample pair according to the matching copy and the reference copy comprises: Extracting second text element information from the matching text and extracting third text element information from the reference text; Constructing text preference information according to the second text element information and the third text element information; The matching copy and the reference copy are used as copy preference labels, and the copy preference information is used as a copy preference sample; A copywriting preference sample pair is constructed based on the copywriting preference label and the copywriting preference sample.

6. The method for generating text according to any one of claims 1 to 5, wherein the optimizing a large language model based on the text generation sample pairs and the text preference sample pairs, and generating a target text corresponding to the target object using the optimized large language model, comprises: Optimizing the large language model according to the copy generation sample pair and the copy preference sample pair until a target large language model that meets the optimization stop condition is obtained; The object description information corresponding to the target object is input into the target large language model for processing to obtain a target text that matches the target object to the user group.

7. According to the copywriting generation method according to any one of claims 1 to 5, the determination of the user group comprises: Determining multiple users associated with the target object; The multiple users are divided according to a preset division rule to obtain at least two initial user groups, and the at least two initial user groups are used as user groups associated with the target object.

8. The method for generating text according to any one of claims 1 to 5, after the step of generating the target text corresponding to the target object by using the optimized large language model is executed, further comprising: When the new copy is obtained, determine the target user group that matches the new copy; The newly added copy is used as a matching copy that matches the target user group, and steps of constructing a copy generation sample pair according to the matching copy and constructing a copy preference sample pair according to the matching copy and the reference copy are performed.

9. A copywriting generation method, comprising: Determine a target product and a user group associated with the target product; Constructing a multi-dimensional push copy set for the target product, selecting a matching copy that matches the user group from the multi-dimensional push copy set, and determining a reference copy from the multi-dimensional push copy set; Constructing a copy generation sample pair according to the matching copy, and constructing a copy preference sample pair according to the matching copy and the reference copy; Optimizing the large language model based on the copy generation sample pair and the copy preference sample pair, and using the optimized large language model to generate a target push copy corresponding to the target product; The target push copy is sent to terminal devices of users included in the user group.

10. A method for generating multimodal information, comprising: Determining a target object and a user group associated with the target object; Constructing a multi-dimensional multi-modal information set for the target object, selecting matching multi-modal information that matches the user group from the multi-dimensional multi-modal information set, and determining reference multi-modal information from the multi-dimensional multi-modal information set; Constructing a multimodal information generation sample pair according to the matching multimodal information, and constructing a multimodal information preference sample pair according to the matching multimodal information and the reference multimodal information; The large language model is optimized based on the multimodal information generation sample pairs and the multimodal information preference sample pairs, and the target multimodal information corresponding to the target object is generated using the optimized large language model.

11. A computing device comprising: Memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the computer-executable instructions are executed by the processor, the steps of the method according to any one of claims 1 to 10 are implemented.

12. A computer-readable storage medium storing computer-executable instructions, wherein the computer-executable instructions, when executed by a processor, implement the steps of the method according to any one of claims 1 to 10.

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

Images