Method and apparatus for generating copy on basis of large language model, and storage medium and computer device

Abstract

Disclosed in the present application are a method and apparatus for generating a copy on the basis of a large language model, and a storage medium and a computer device. The method comprises: on the basis of a target copy generation requirement, acquiring at least one data sample corresponding to at least one historical user, wherein each data sample comprises various types of user feature data of the corresponding historical user, user browsing data of the corresponding historical user and user interaction data of the corresponding historical user, and a browsed object in the user browsing data matches the target copy generation requirement; using data samples, to which user interaction data that represents the occurrence of interaction belong, to construct positive samples, and using remaining data samples to construct negative samples; training a decision tree by means of the positive samples and the negative samples, and respectively constructing user feature sets on the basis of branches of the decision tree; and on the basis of the user feature sets and the target copy generation requirement, constructing copy generation prompt information, and on the basis of the copy generation prompt information, generating a target copy by means of a large language model. By means of the present application, copies that match preferences of different user groups can be rapidly generated.

Description

Copywriting generation method and device based on large language model, storage medium and computer device

[0001] The present application claims priority to the Chinese patent application No. 202410846000.7, filed on June 27, 2024, and entitled "Copywriting generation method and device based on large language model, storage medium and computer device", the whole content of which is incorporated herein by reference. TECHNICAL FIELD

[0002] The present application relates to the technical field of natural language processing, in particular to a copywriting generation method and device based on a large language model, a storage medium and a computer device. BACKGROUND

[0003] It has always been one of the core strategies for different user groups to develop differentiated copywriting in various fields to achieve information recommendation. Compared with fixed and unified copywriting, tailor-made copywriting for different user groups can more accurately grasp the interest points of different user groups and help meet the needs of different user groups.

[0004] Traditional copywriting generation processes usually rely on the experience and intuition of professional copywriters, which is not only inefficient but also difficult to meet the growing personalized needs. With the development of artificial intelligence technology, especially the emergence of large language models, automated copywriting has become feasible. However, existing automated copywriting methods are still insufficient to understand complex user characteristics, resulting in copywriting that often lacks pertinence and personalization.

[0005] SUMMARY

[0006] Therefore, the embodiments of the present application provide a copywriting generation method and device based on a large language model, a storage medium and a computer device, so that the generated copywriting can more accurately match the preferences and needs of different user groups, improve the personalization degree and relevance of copywriting, better attract the attention of users, help improve the interaction willingness and satisfaction of users, and thus improve the user experience.

[0007] According to one aspect of the present application, a copywriting generation method based on a large language model is provided, the method comprising:

[0008] According to the target copywriting generation requirement, at least one historical user corresponding to at least one data sample is obtained, wherein each data sample includes a plurality of user feature data, user browsing data and user interaction data of the corresponding historical user, the browsed object in the user browsing data matches the target copywriting generation requirement, and the user interaction data indicates whether to interact with the browsed object;

[0009] construct positive samples by using data samples to which user interaction data representing interactions belong, and construct negative samples by using remaining data samples;

[0010] train a decision tree by using the positive samples and the negative samples, and respectively construct user feature sets based on branches of the decision tree, wherein each user feature set includes at least one user feature type label;

[0011] For any user feature set, generate copy generation prompt information according to the user feature set and the target copy, and generate a target copy based on the copy generation prompt information through a large language model.

[0012] Optionally, before the constructing positive samples by using data samples to which user interaction data representing interactions belong, and the constructing negative samples by using remaining data samples, the method comprises:

[0013] perform feature binning on continuous features in the user feature data to obtain a plurality of feature bins corresponding to each continuous feature;

[0014] use discrete features in the user feature data and the feature bins as training features, and perform one-hot encoding on each piece of user feature data based on the training features to obtain user encoding features corresponding to each piece of user feature data;

[0015] Correspondingly, the constructing positive samples by using data samples to which user interaction data representing interactions belong, and the constructing negative samples by using remaining data samples, comprise:

[0016] construct positive samples by using user encoding features corresponding to data samples to which user interaction data representing interactions belong, and construct negative samples by using user encoding features corresponding to remaining data samples.

[0017] Optionally, the training a decision tree by using the positive samples and the negative samples, and the respectively constructing user feature sets based on branches of the decision tree, comprise:

[0018] train a decision tree based on a preset feature importance condition by using the positive samples and the negative samples, wherein the decision tree includes a plurality of branches, each branch includes one leaf node and at least one decision node, the user feature types represented by each decision node all meet the preset feature importance condition, and each leaf node is marked with a classification label of the leaf node;

[0019] determine target branches in the branches according to the classification labels of the leaf nodes, and respectively construct user feature sets based on the user feature types represented by each decision node in each target branch.

[0020] Optionally, each leaf node is also marked with the number of samples split from the root node to the leaf node; based on the user feature types represented by each decision node in each target branch, user feature sets are respectively constructed, including:

[0021] A proportion of the number of samples corresponding to the leaf nodes in the target branch to the total number of data samples is determined, and the target branch is screened based on the proportion;

[0022] User feature sets are respectively constructed using the user feature types represented by each decision node in each screened target branch.

[0023] Optionally, after the target script is generated by the large language model based on the script generation prompt information, the method further comprises:

[0024] In the case where the target script touch display signal based on the target user is detected, a target display script is determined from the target scripts corresponding to each user feature set based on the target user features corresponding to the target user;

[0025] The target display script is reached to the target user.

[0026] Optionally, in the case where the target script touch display signal based on the target user is detected, a target display script is determined from the target scripts corresponding to each user feature set based on the target user features corresponding to the target user, comprising:

[0027] In the case where the target script touch display signal based on the target user is detected, a target user feature set matching the target user features is queried from the user feature sets based on the target user features corresponding to the target user, and a target script associated with the target user feature set is determined as the target display script;

[0028] If the user feature sets do not match the target user features, the standard script corresponding to the target script generation requirement is taken as the target display script.

[0029] Optionally, after the target display script is reached to the target user, the method further comprises:

[0030] User behavior data after each target script is reached to the user is obtained, and script effects of each target script are respectively determined according to the user behavior data corresponding to each target script;

[0031] The target script that does not meet the preset script effect in the target script is determined according to the script effect, and the target script is adjusted and replaced.

[0032] Optionally, the adjusting and replacing of the to-be-adjusted script includes:

[0033] According to the user feature set corresponding to the to-be-adjusted script, the target script generation requirement, and the adjustment requirement corresponding to the to-be-adjusted script, a script adjustment prompt information is constructed, and a large language model is used to generate an adjusted script based on the script adjustment prompt information to replace the to-be-adjusted script.

[0034] Optionally, the adjusting and replacing of the to-be-adjusted script includes:

[0035] The user feature set corresponding to the to-be-adjusted script is taken as a to-be-adjusted set, the user behavior data corresponding to the to-be-adjusted script is taken as a retraining sample, and a new decision tree is trained using the retraining sample, and a new user feature set is constructed based on each branch of the new decision tree to replace the to-be-adjusted set.

[0036] For any new user feature set, a new script generation prompt information is constructed according to the new user feature set and the target script generation requirement, and a large language model is used to generate a new target script based on the new script generation prompt information to replace the original target script corresponding to the to-be-adjusted set.

[0037] Optionally, the target script generation requirement includes at least one of the following: a target object, a target object category, at least one object under a target theme, and at least one object category under a target theme, wherein the object includes a commodity or at least one multimedia material.

[0038] According to another aspect of the present application, a script generation device based on a large language model is provided, and the device includes:

[0039] A data processing module is configured to obtain at least one data sample corresponding to at least one historical user according to a target script generation requirement, wherein each data sample includes a plurality of user feature data, user browsing data, and user interaction data of a corresponding historical user, the browsed object in the user browsing data matches the target script generation requirement, and the user interaction data indicates whether an interaction is generated on the browsed object.

[0040] A model training module is configured to construct a positive sample using a data sample in which the user interaction data indicating that an interaction is generated, and construct a negative sample using the remaining data samples; and train a decision tree using the positive sample and the negative sample.

[0041] A model analysis module is configured to construct a user feature set based on each branch of the decision tree, wherein each user feature set includes at least one user feature type label.

[0042] The advertising copy generation module is configured to, for any user feature set, construct an advertising copy generation prompt according to the user feature set and the target advertising copy generation requirement, and generate a target advertising copy based on the advertising copy generation prompt by using a large language model.

[0043] Optionally, the data processing module is further configured to:

[0044] The continuous features in the user feature data are subjected to feature bucketing, so as to obtain a plurality of feature buckets corresponding to each continuous feature.

[0045] The discrete features in the user feature data and the feature buckets are taken as training features, and each piece of user feature data is subjected to one-hot encoding based on the training features, so as to obtain user encoding features corresponding to each piece of user feature data.

[0046] Correspondingly, the model training module is further configured to:

[0047] The positive samples are constructed by using the user encoding features corresponding to the data samples representing the user interaction data, and the negative samples are constructed by using the user encoding features corresponding to the remaining data samples.

[0048] Optionally, the model training module is further configured to train a decision tree based on the positive samples and the negative samples based on a preset feature importance condition, wherein the decision tree comprises a plurality of branches, each branch comprises a leaf node and at least one decision node, the user feature types represented by each decision node all meet the preset feature importance condition, and each leaf node is marked with a classification label of the leaf node.

[0049] The model analysis module is further configured to determine a target branch in each branch according to the classification label of the leaf node, and construct a user feature set based on the user feature types represented by each decision node in each target branch.

[0050] Optionally, each leaf node is further marked with a sample quantity from a root node to the leaf node; and the model analysis module is further configured to:

[0051] determine a proportion of the sample quantity corresponding to the leaf node in the target branch in a total sample quantity of the data samples, and screen the target branch based on the proportion.

[0052] The user feature sets are constructed based on the user feature types represented by each decision node in each screened target branch.

[0053] Optionally, the apparatus further comprises an advertising copy display module configured to:

[0054] In a case where the target user-based target advertisement reach display signal is detected, a target advertisement reach display is determined in a target advertisement corresponding to each user feature set based on a target user feature corresponding to the target user.

[0055] The target advertisement reach display is reached to the target user.

[0056] Optionally, the advertisement display module is further configured to:

[0057] In a case where the target user-based target advertisement reach display signal is detected, a target advertisement reach display is determined in a target advertisement corresponding to each user feature set based on a target user feature corresponding to the target user.

[0058] If the user feature sets are all not matched with the target user feature, a standard advertisement corresponding to the target advertisement generation requirement is taken as the target advertisement reach display.

[0059] Optionally, the device further comprises an evaluation and optimization module configured to:

[0060] User behavior data after each target advertisement reach display is obtained, and an advertisement effect of each target advertisement is determined based on user behavior data corresponding to each target advertisement.

[0061] A to-be-adjusted advertisement that does not meet a preset advertisement effect in the target advertisement is determined based on the advertisement effect, and the to-be-adjusted advertisement is adjusted and replaced.

[0062] Optionally, the evaluation and optimization module is further configured to:

[0063] Based on a user feature set corresponding to the to-be-adjusted advertisement, the target advertisement generation requirement, and an adjustment requirement corresponding to the to-be-adjusted advertisement, an advertisement adjustment prompt information is constructed, and an adjusted advertisement is generated by a large language model based on the advertisement adjustment prompt information to replace the to-be-adjusted advertisement.

[0064] Optionally, the evaluation and optimization module is further configured to: take the user feature set corresponding to the to-be-adjusted advertisement as a to-be-adjusted set, obtain user behavior data corresponding to the to-be-adjusted advertisement as a retraining sample, and train a new decision tree by using the retraining sample, and based on each branch of the new decision tree, a new user feature set is constructed to replace the to-be-adjusted set.

[0065] For any new user feature set, new copy generation prompt information is constructed according to the new user feature set and the target copy generation requirement, and a new target copy is generated by a large language model based on the new copy generation prompt information to replace the original target copy corresponding to the to-be-adjusted set.

[0066] Optionally, the target copy generation requirement includes at least one of the following: a target object, a target object category, at least one object under a target theme, and at least one object category under a target theme, wherein the object includes a commodity or at least one multimedia material.

[0067] According to another aspect of the present application, a storage medium having a computer program stored thereon is provided, and the program is executed by a processor to implement the above-mentioned copy generation method based on a large language model.

[0068] According to another aspect of the present application, a computer device is provided, which includes a storage medium, a processor, and a computer program stored on the storage medium and executable on the processor, and the processor implements the above-mentioned copy generation method based on a large language model when executing the program.

[0069] By means of the above technical solution, the copy generation method and device based on a large language model, the storage medium, and the computer device provided by the embodiments of the present application filter user browsing data of historical users according to target copy generation requirements, and further obtain user interaction data corresponding to the user browsing data and user feature data of the historical users to jointly construct data samples. Based on whether the user interaction data in the data samples indicates that the user has an interaction behavior on the browsed object, the data samples are divided into positive samples and negative samples, so as to train a decision tree. The user feature set is constructed based on each branch of the decision tree, so as to realize the combination of user features. Finally, prompt information is constructed for each user feature set, and a target copy of each user feature set is generated by a large language model, thereby forming differentiated copies of different user feature groups. The embodiments of the present application train a decision tree by using various feature data and user behavior data of historical users, and construct a user feature set by analyzing user features corresponding to each branch of the decision tree, so as to generate a copy for different user feature sets. The generated copy can more accurately match the preferences and requirements of different user groups, improve the individuality and relevance of the copy, better attract the attention of users, help to improve the interaction willingness and satisfaction of users, and thus improve the user experience.

[0070] The above description is only a summary of the technical solutions of the present application. In order to more clearly understand the technical means of the present application, the embodiments of the present application can be implemented in accordance with the content of the description, and in order to make the above and other purposes, features and advantages of the present application more obvious and easy to understand, the following specific embodiments of the present application are described. BRIEF DESCRIPTION OF DRAWINGS

[0071] The accompanying drawings, which are included to provide a further understanding of the application and are incorporated in and constitute a part of this application, illustrate embodiments of the application and serve to explain the principles of the application. In the drawings:

[0072] FIG. 1 shows a flowchart of a large language model-based copywriting method according to an embodiment of the present application;

[0073] FIG. 2 shows a flowchart of another large language model-based copywriting method according to an embodiment of the present application;

[0074] FIG. 3 shows a structural diagram of a large language model-based copywriting device according to an embodiment of the present application;

[0075] FIG. 4 shows a structural diagram of a large language model-based copywriting device according to an embodiment of the present application. DETAILED DESCRIPTION

[0076] Hereinafter, the present application will be described in detail with reference to the accompanying drawings and embodiments. It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict.

[0077] In the present embodiment, a large language model-based copywriting method is provided, as shown in FIG. 1, which includes:

[0078] In step 101, at least one data sample corresponding to at least one historical user is obtained according to a target copywriting generation requirement, wherein each data sample includes a plurality of user feature data, user browsing data and user interaction data of a corresponding historical user, the browsed object in the user browsing data matches the target copywriting generation requirement, and the user interaction data indicates whether an interaction is generated on the browsed object.

[0079] In the embodiments of the present application, for the target copywriting generation requirement, data samples for decision tree training are obtained, which are generally data generated based on the real behavior of historical users, wherein the data samples contain user browsing data, user interaction data and user feature data. Whether the behavior data of the user will be used as a data sample depends on whether the browsed object matches the target copywriting generation requirement. If it matches, a data sample can be determined based on the user's browsing behavior of the browsed object. For example, the target copywriting generation requirement is generated for a crayfish theme activity, and the behavior data of the historical user browsing crayfish commodities can be used to form a data sample. After the user browses the crayfish commodities, whether the user generates an interactive behavior on the crayfish commodities can reflect the user's interest in the crayfish commodities, and the user feature data is conducive to inducing which user with what characteristics is interested in or not interested in the browsed object.

[0080] Optionally, the target copywriting generation requirement includes at least one of the following: a target object, a target object category, at least one object under a target theme, and at least one object category under a target theme, wherein the object includes a commodity or at least one multimedia material.

[0081] Taking an e-commerce platform as an example, the target copywriting generation requirement can be a requirement formulated for the purpose of a marketing activity, a theme activity or the like launched by the platform. For example, for a "late-night snack" theme activity of a take-out platform or a local life platform, the target copywriting generation requirement can include the late-night snack food category, and the data sample can be determined according to the user's behavior of browsing the late-night snack food category on the platform. The data sample can include user browsing data (indicating that the historical user browsed a certain commodity of the late-night snack category), user interaction data (indicating whether the historical user generates a click, collection, order, sharing or the like on the commodity of the late-night snack category browsed by the user), and a plurality of user feature data of the historical user, wherein the user feature data can include user basic portrait data (such as age, gender, consumption level, category preference, etc.), user behavior portrait data (such as the purchase frequency data of the user on various commodities under the late-night snack category, etc.).

[0082] Taking a multimedia material sharing platform as an example, the target copy generation requirement can be a requirement formulated for the purpose of a column activity, a promotion activity, or the like launched by the platform. For example, for a theme activity of detective reasoning novels of the book sharing platform, the target copy generation requirement can include detective reasoning novels, and the data sample can be data determined according to the behavior of the user browsing a book introduction list containing detective reasoning novels on the platform. The data sample can include user browsing data (indicating that the historical user has browsed detective reasoning novels), user interaction data (indicating whether the user has generated a click, collection, or the like interaction behavior on the detective reasoning novels browsed by the user), and various user feature data of the historical user.

[0083] In step 102, positive samples are constructed by using data samples to which user interaction data indicating generation of interaction belongs, and negative samples are constructed by using remaining data samples.

[0084] In the embodiments of the present application, after collecting the data samples, the data samples can be classified. Data samples related to the target copy generation requirement and having interaction behaviors (such as clicking, purchasing, and the like) are marked as positive samples, indicating that these samples are related to the target requirement. The remaining data samples are marked as negative samples, indicating that these samples are not related or have low relevance to the target requirement.

[0085] In the embodiments of the present application, optionally, before step 102, feature binning can be performed on continuous features in the user feature data to obtain a plurality of feature bins corresponding to each continuous feature. Discrete features in the user feature data and the feature bins are used as training features, and each piece of user feature data is one-hot encoded based on the training features to obtain user encoding features corresponding to each piece of user feature data. Correspondingly, step 102 can include: constructing positive samples by using user encoding features corresponding to data samples to which user interaction data indicating generation of interaction belongs, and constructing negative samples by using user encoding features corresponding to remaining data samples.

[0086] In the above embodiments, for continuous features in the user feature data (such as age, etc.), the values of these features are continuous and can have a large number of unique values. In order to simplify the model and improve performance, these continuous features are usually subjected to bucketing processing. Feature bucketing is to divide the value range of a continuous feature into several intervals (i.e. "buckets"), and then replace the original continuous value with the identification of the bucket to which it belongs. In this way, the continuous feature is converted into a discrete feature. In machine learning, discrete features (including continuous features subjected to bucketing processing and features that are originally discrete) usually need to be converted into a format suitable for model processing. One-hot encoding is a method of converting discrete features into binary vectors. For each feature, its possible values will have a corresponding binary bit. If the feature takes a certain value, the corresponding binary bit is 1, and the other bits are 0. In this process, the discrete features in the user feature data and the continuous features subjected to bucketing processing (i.e. feature buckets) are converted into one-hot encoded form to form user encoded features. Further, for the encoded data samples, the data samples related to user interaction data (such as clicks, purchases, etc.) can be considered to be positively correlated with the target text generation requirement and can be labeled as positive samples. The remaining data samples, i.e. data samples without interaction, are considered to be less relevant or irrelevant to the target task, and are therefore labeled as negative samples. The embodiments of the present application can simplify the data through feature bucketing and one-hot encoding, making it easier for the model to learn. And for some features, there can be missing values. Through one-hot encoding, missing values can be treated as a special value, so that they are explicitly handled in the model.

[0087] In specific application scenarios, after collecting data samples, the staff also screen the user feature types in combination with the activity appeal to avoid the problem of high training cost and low efficiency caused by using all user features. For example, for the floor theme copy generation requirement, the age, gender, category preference, and purchase pen number labels can be screened. Further, the user feature data in the data samples can be encoded, and the One-Hot encoding (one-hot encoding) method can be used. Among them, the numerical discrete class feature is directly subjected to one-hot encoding to generate a new feature column. For example, for the user_status (user identity) feature, the values are new user / old user, etc., and user_status = old user is encoded into a new feature column, such as user_status_is_old. When the user is an old user, this column takes the value 1, otherwise 0. The continuous value feature is processed by combining equal-width / equal-frequency bucketing with one-hot encoding, such as age, which is first bucketed into 36-40, and then one-hot encoded to obtain the feature "tb_age_36_40" in the column.

[0088] In step 103, the decision tree is trained by using the positive samples and the negative samples, and user feature sets are respectively constructed based on branches of the decision tree, wherein each user feature set includes at least one user feature type label.

[0089] In the embodiments of the present application, a decision tree model is trained by using positive samples and negative samples, for example, a GBDT decision tree model is selected. And each branch of the trained decision tree is divided based on one or more user features, forming a plurality of user feature sets. The user features in these sets can divide different users into a plurality of groups based on features, and these groups are divided based on the target script generation requirements.

[0090] In the embodiments of the present application, step 103 optionally includes:

[0091] In step 103-1, the decision tree is trained based on a preset feature importance condition by using the positive samples and the negative samples, wherein the decision tree includes a plurality of branches, each branch includes a leaf node and at least one decision node, the user feature types represented by each decision node meet the preset feature importance condition, and each leaf node is marked with a classification label of the leaf node.

[0092] In step 103-2, a target branch in each branch is determined according to the classification label of the leaf node, and user feature sets are respectively constructed based on the user feature types represented by each decision node in each target branch.

[0093] In this embodiment, one or more feature importance conditions are set before starting to train the decision tree. These conditions are usually based on activity requirements, data characteristics or model performance considerations, and are used to filter out features that should be focused on in the decision tree. The decision tree is trained using positive samples (i.e. data samples related to user interaction data) and negative samples (i.e. data samples that do not generate interactions). During the training process, according to the preset feature importance conditions, user features that meet these conditions are selected as decision nodes, and features that are not significantly important or not important are filtered out. For example, a threshold is set to filter out effective label features with a feature importance greater than 0.01. By recursively selecting optimal features and splitting points, the entire decision tree structure is constructed. The decision tree is composed of multiple branches, each branch starts from the root node and passes through a series of decision nodes to reach the leaf node, each decision node represents a user feature and distributes samples to different sub-branches according to the value of the feature, and the leaf node is the end point of the decision tree, marked with the classification label corresponding to the leaf node (here, the label indicating positive samples or negative samples). After the decision tree is trained, the classification label of the leaf node (i.e. positive samples or negative samples) is used to determine which branches are "target branches". For each target branch, all decision nodes in the branch are traversed. Each decision node represents a user feature type. Collecting these user feature types forms a user feature set. Since the paths of each target branch are different, different target branches will construct different user feature sets. Based on the user feature sets constructed by different target branches, guidance can be provided for subsequent copy generation. For example, based on the features of a specific user feature set, a copy can be generated that is more in line with the preferences of the user group.

[0094] In the embodiments of the present application, each leaf node is also marked with the number of samples that split from the root node to the leaf node; step 103-2 includes: determining the proportion of the number of samples corresponding to the leaf node in the target branch to the total number of data samples, and filtering the target branch based on the proportion; and constructing a user feature set using the user feature types represented by each decision node in each filtered target branch.

[0095] In the above embodiments, in the decision tree, each leaf node is not only marked with a classification label (such as a positive sample or a negative sample), but also can be marked with the number of samples from the root node to the leaf node. This number reflects the size of the sample set represented by the leaf node, and further can reflect the proportion of the user group represented by the leaf node in the entire data set, so as to filter the target branch based on the number of samples. First, determine the proportion of the number of samples corresponding to the leaf node in each target branch (i.e. the branch in which the leaf node containing the positive sample or the negative sample is located) in the total number of samples in the entire data sample. Based on the proportion calculated above, the target branch can be filtered. For example, a threshold can be set to only keep those target branches whose proportion exceeds the threshold. In this way, it can be ensured that attention is focused on those target branches containing enough samples, while branches with fewer samples, which may not be typical, are ignored. After the above filtering, the user feature set is constructed using the user feature types represented by each decision node in each filtered target branch. These user feature sets represent user feature combinations that are more relevant to the target copy generation demand. The embodiments of the present application can ensure that the user feature set constructed is based on a representative user group with sufficient sample quantity through filtering based on the number of samples of the leaf node. This helps to improve the generalization ability of the feature set and the effect in actual application. And if all target branches are considered without filtering, many user feature sets may be obtained, some of which may not have practical significance or representativeness, while filtering can reduce unnecessary computational cost and ensure that only those truly important user feature sets are focused on. In addition, the filtered user feature set is more concise and clear, which helps to better understand which user features have an important impact on the target task, thereby enhancing the explainability of the model.

[0096] Step 104, for any user feature set, constructing a copy generation prompt information according to the user feature set and the target copy generation demand, and generating a target copy through a large language model based on the copy generation prompt information.

[0097] In the embodiments of the present application, for each user feature set, the target copywriting generation requirements and the user feature labels contained in the set are combined to construct copywriting generation prompt information. The target copywriting generation requirements can also include activity-related information, and the copywriting generation prompt information can also include the corresponding copywriting style determined according to the user feature set, so that the generated copywriting is closer to the style that the corresponding user group likes. Then, a large language model (such as GPT, BERT, etc.) or a multi-modal model is used to generate target copywriting based on these prompt information. Thus, differentiated target copywriting for different user groups is accurately formed, and highly relevant and personalized content to user group characteristics is provided to meet the unique needs of different user groups. In addition, for the large language model to generate target copywriting, further checking and optimization can be performed, which can specifically check the semantic consistency, relevance, creativity, etc. of the generated target copywriting. For example, semantic consistency checking can be performed by matching the semantics of the standard copywriting and the target copywriting, relevance checking can be performed based on pre-set artificial rules, and creativity of the target copywriting can be artificially evaluated by sampling. When it is checked that there is a deficiency in some aspect, the problems that need to be improved, the target copywriting generation requirements, the target copywriting that needs to be improved, and the user feature set are used to generate prompt information again, and the target copywriting is generated again by the large language model. Artificial optimization can also be performed when it is found that the target copywriting has deficiencies, which is not limited herein.

[0098] In one example, for the user feature set (25-34 years old, urban white-collar, healthy meal lover), the individualized prompt template for night snack small lobster is designed as follows: “Considering that you are an urban white-collar between the ages of 25 and 34, and have a special preference for healthy meals, we want to recommend a night snack option that meets your lifestyle: healthy small lobster. Given your busy work requirements during the day, high standards for food quality, and pursuit of life quality, please write a marketing copy for this type of customer, focusing on the three aspects of health, convenience, and deliciousness, while incorporating the life scenes of urban white-collar workers to enhance their interest in choosing our night snack small lobster.”

[0099] By applying the technical solution of the embodiment, the user browsing data of the historical user is filtered according to the target text generation requirement, and the user interaction data corresponding to the user browsing data and the user feature data of the historical user are further obtained to jointly construct a data sample. Based on whether the user interaction data in the data sample indicates that the user has an interaction behavior on the browsed object, the data sample is divided into positive samples and negative samples, thereby training a decision tree. The user feature set is constructed based on each branch of the decision tree to realize the combination of user features. Finally, the prompt information is constructed for each user feature set, the target text of each user feature set is generated by the large language model, and the differentiated text of different user feature groups is formed. The embodiment of the application trains the decision tree by using the various feature data and user behavior data of the historical user, and constructs the user feature set by analyzing the user features corresponding to each branch of the decision tree, thereby generating the text for different user feature sets. The generated text can more accurately match the preferences and requirements of different user groups, improve the personalization degree and relevance of the text, better attract the attention of users, help to improve the interaction willingness and satisfaction of users, and thereby improve the user experience.

[0100] Further, as a refinement and expansion of the above embodiment, in order to completely describe the specific implementation process of the embodiment, another text generation method based on a large language model is provided, as shown in FIG. 2. The method includes:

[0101] Step 201, in the case where a target text triggering display signal based on a target user is detected, a target display text is determined in the target text corresponding to each user feature set based on the target user feature corresponding to the target user.

[0102] In the embodiment of the application, when a target text triggering display signal based on a target user (any user) is detected (for example, the user visits a certain page or clicks a certain button), a target display text most suitable for the user can be selected and determined from the target texts corresponding to a plurality of user feature sets prepared in advance according to the target user feature corresponding to the target user. In addition, the target text and the user feature set can be associated and stored in advance, and can be directly read when online display is required.

[0103] Optionally, step 201 includes: in the case where a target text triggering display signal based on a target user is detected, a target user feature set matching the target user feature is queried from the user feature set based on the target user feature corresponding to the target user, and a target text associated with the target user feature set is determined as the target display text; if the user feature set does not match the target user feature, a standard text corresponding to the target text generation requirement is taken as the target display text.

[0104] In the above embodiment, when the system detects a target display signal based on the target user's target script, this signal can be triggered by a certain behavior or event of the user, such as the user accessing a certain page, clicking a certain button, etc. The target user's target user features are obtained, and the pre-defined user feature set is queried to determine whether there is a user feature set matching the target user features. If a user feature set matching the target user features is found, since each target script is designed and optimized according to a user group similar to the target user features, it is more likely to arouse the interest and response of the corresponding group of users, so the target script associated with the set can be directly determined as the target display script. If no matching set is found among all user feature sets (that is, the target user's features do not have a direct corresponding classification in the existing set), an alternative solution can be adopted. This alternative solution is to use the standard script corresponding to the target script generation requirement as the target display script. The standard script is usually suitable for most users or users without explicit feature classification. Although it may not be as accurate as the script matching the specific user features, it can still provide some attraction and effect. Through this embodiment, both customized display scripts according to the user's personalized features and standard scripts to ensure basic display effects in the absence of direct matching features can be provided, which helps to improve user experience and script display effect.

[0105] Step 202, the target display script is reached to the target user.

[0106] In the embodiments of the present application, after determining the target display script, it is reached to the target user in a suitable way (such as pop-up window, pop-up page in APP, email, SMS, APP push, etc.).

[0107] Step 203, the user behavior data after each target script reaches the user is obtained, and the script effect of each target script is determined according to the user behavior data corresponding to each target script.

[0108] In the embodiments of the present application, after the target script reaches the user, the system can collect the user's subsequent behavior data, which usually includes whether the user clicks the script, whether the user makes a purchase after clicking, whether the user views more content, etc. According to these user behavior data, the system can evaluate the script effect of each target script to evaluate the influence degree of each target script on user behavior. For example, the user behavior data includes the user's clicking on the script, the user's clicking on the goods after entering the goods list (or multimedia material list), whether the user places an order after clicking, whether the user collects, etc. The script effect of the target script is evaluated comprehensively.

[0109] In step 204, the text to be adjusted that does not meet the preset text effect in the target text is determined according to the text effect, and the text to be adjusted is adjusted and replaced.

[0110] In the embodiments of the present application, based on the evaluation result of the text effect, it can be determined which text does not meet the preset text effect standard (such as too low click rate, insufficient conversion rate, etc.), and these texts are called "text to be adjusted". For the text to be adjusted, the system will further optimize and replace. Optimization may include modifying the text content, modifying the user feature set corresponding to the text and regenerating the text content. In order to more accurately understand the user's needs and preferences, optimize the user experience.

[0111] In the embodiments of the present application, optionally, the adjustment and replacement of the text to be adjusted in step 204 include: constructing a text adjustment prompt information according to the user feature set corresponding to the text to be adjusted, the target text generation requirement and the adjustment requirement corresponding to the text to be adjusted, and generating an adjusted text based on the text adjustment prompt information through a large language model to replace the text to be adjusted.

[0112] In this embodiment, when a certain text is determined to be a text to be adjusted (i.e. the text effect of the text does not meet the preset standard), the adjustment requirement of this text needs to be determined first. For example, the adjustment requirement can include aspects in which the text performs poorly (such as low click rate, poor conversion rate, etc.), and can also include the reasons for the poor text effect. Or the adjustment requirement can not contain the above information and only reflect "please regenerate the text". Based on the user feature set corresponding to the text to be adjusted, combined with the generation requirement of the target text (such as the desired user engagement, purchase intention, etc.), and the adjustment requirement of the text to be adjusted (such as the part that needs to be modified, the desired improvement effect, etc.), a text adjustment prompt information is constructed. The text adjustment prompt information constructed above is input into a large language model (such as GPT, BERT, etc.). The large language model will automatically generate one or more adjusted texts according to the input prompt information. These texts will try to solve the problems in the text to be adjusted, and try to meet the generation requirement of the target text as much as possible. Finally, the adjusted text generated by the large language model is replaced by the original text to be adjusted. This replacement process may need to be evaluated and screened to ensure that the new text not only solves the original problem, but also is better than the original text in overall effect. The embodiments of the present application automate and intelligentize the generation and optimization process of the text. By combining the user feature set, the target text generation requirement and the adjustment requirement of the text to be adjusted, a targeted text adjustment prompt information can be constructed. Then, by using the powerful generation capability of the large language model, the required adjusted text can be quickly generated, thereby improving the efficiency and quality of text optimization.

[0113] In the embodiment of the present application, the adjusting and replacing of the to-be-adjusted script in step 204 comprises: taking the user feature set corresponding to the to-be-adjusted script as a to-be-adjusted set, obtaining user behavior data corresponding to the to-be-adjusted script as a retraining sample, and training a new decision tree using the retraining sample, and constructing a new user feature set based on each branch of the new decision tree to replace the to-be-adjusted set; for any new user feature set, generating a new script generation prompt information according to the new user feature set and the target script generation requirement, and generating a new target script based on the new script generation prompt information through a large language model to replace the original target script corresponding to the to-be-adjusted set.

[0114] In this embodiment, first, the user feature set corresponding to the to-be-adjusted script is regarded as a to-be-adjusted set. Then, the user behavior data related to the to-be-adjusted script is obtained as a retraining sample. These behavior data reflect the user's interaction with the current script and the user's participation in the activity corresponding to the current script, which is crucial for subsequent model training. The retraining sample can be the same as the data sample structure described above. A new decision tree model is trained using the obtained retraining sample. Based on each branch of the new decision tree, a new user feature set is constructed. These new sets will update and optimize the user's classification and features according to the new decision tree model. For each new user feature set, a new script generation prompt information is constructed in combination with the target script generation requirement (such as the desired user participation, purchase intention, etc.). This prompt information will serve as a guide for the large language model to generate a new script, containing specific requirements for target users, target effects, and script style. Using a large language model, a new target script is generated according to the new script generation prompt information. The large language model can learn the rules and patterns of language based on a large amount of text data and generate text that conforms to these rules and patterns. The generated new target script is used to replace the original target script corresponding to the to-be-adjusted set. This process ensures that the update and optimization of the script are based on the latest behavior data of the user and the training results of the decision tree model. The way to retrain the decision tree can be consistent with the way described in the embodiment of FIG. 1, which will not be repeated here. By training a new decision tree model using user behavior data, the actual needs and interests of users can be more accurately captured to optimize and reconstruct the user feature set; and using a large language model to generate a new target script can ensure the language expression and content quality of the script. The entire process not only improves the relevance and effectiveness of the script, but also improves the efficiency and quality of script generation.

[0115] Further, as a specific implementation of the method of FIG. 1, the embodiment of the present application provides a script generation device based on a large language model, as shown in FIG. 3, which comprises:

[0116] a data processing module configured to obtain at least one data sample corresponding to at least one historical user according to a target script generation requirement, wherein each data sample comprises a plurality of user feature data, user browsing data and user interaction data of a corresponding historical user, the browsed object in the user browsing data matches the target script generation requirement, and the user interaction data indicates whether an interaction is generated on the browsed object;

[0117] a model training module configured to construct positive samples by using data samples in which the user interaction data indicating that an interaction is generated, and construct negative samples by using the remaining data samples, and train a decision tree by using the positive samples and the negative samples;

[0118] a model analysis module configured to construct a user feature set based on each branch of the decision tree, wherein each user feature set comprises at least one user feature type label;

[0119] a script generation module configured to, for any user feature set, construct script generation prompt information according to the user feature set and the target script generation requirement, and generate a target script based on the script generation prompt information by using a large language model.

[0120] Optionally, the data processing module is further configured to:

[0121] perform feature bucketing on continuous features in the user feature data to obtain a plurality of feature buckets corresponding to each continuous feature;

[0122] use discrete features in the user feature data and the feature buckets as training features, and perform one-hot encoding on each piece of user feature data based on the training features to obtain user encoding features corresponding to each piece of user feature data;

[0123] Correspondingly, the model training module is further configured to:

[0124] construct positive samples by using user encoding features corresponding to data samples in which the user interaction data indicating that an interaction is generated, and construct negative samples by using user encoding features corresponding to the remaining data samples.

[0125] Optionally, the model training module is further configured to train the decision tree by using the positive samples and the negative samples based on a preset feature importance condition, wherein the decision tree comprises a plurality of branches, each branch comprises one leaf node and at least one decision node, the user feature types represented by each decision node all meet the preset feature importance condition, and each leaf node is marked with a classification label of the leaf node;

[0126] The model analysis module is further configured to determine a target branch in each branch according to the classification label of the leaf node, and construct a user feature set based on a user feature type represented by each decision node in each target branch.

[0127] Optionally, each leaf node is further marked with a number of samples split from the root node to the leaf node; the model analysis module is further configured to:

[0128] determine a proportion of a number of samples corresponding to the leaf node in the target branch in a total number of the data samples, and filter the target branch based on the proportion;

[0129] construct a user feature set based on a user feature type represented by each decision node in each filtered target branch.

[0130] Optionally, the apparatus further comprises a script display module configured to:

[0131] in a case where a target script reach display signal based on a target user is detected, determine a target display script from target scripts corresponding to each user feature set based on a target user feature corresponding to the target user;

[0132] reach the target display script to the target user.

[0133] Optionally, the script display module is further configured to:

[0134] in a case where a target script reach display signal based on a target user is detected, query a target user feature set matching the target user feature from the user feature set based on a target user feature corresponding to the target user, and determine a target script associated with the target user feature set as the target display script;

[0135] if the user feature set does not match the target user feature, a standard script corresponding to a script generation requirement is taken as the target display script.

[0136] Optionally, as shown in FIG. 4, the apparatus further comprises an evaluation and optimization module configured to:

[0137] obtain user behavior data after each target script reaches a user, and determine a script effect of each target script based on user behavior data corresponding to each target script;

[0138] determine a to-be-adjusted script that does not meet a preset script effect from the target scripts based on the script effect, and adjust and replace the to-be-adjusted script.

[0139] Optionally, the evaluation and optimization module is further configured to:

[0140] According to the user feature set corresponding to the to-be-adjusted script, the target script generation requirement, and the adjustment requirement corresponding to the to-be-adjusted script, a script adjustment prompt information is constructed, and a large language model is used to generate an adjusted script based on the script adjustment prompt information to replace the to-be-adjusted script.

[0141] Optionally, the evaluation optimization module is further configured to: take the user feature set corresponding to the to-be-adjusted script as a to-be-adjusted set, obtain user behavior data corresponding to the to-be-adjusted script as a retraining sample, train a new decision tree by using the retraining sample, and construct a new user feature set based on each branch of the new decision tree to replace the to-be-adjusted set.

[0142] For any new user feature set, a new script generation prompt information is constructed according to the new user feature set and the target script generation requirement, and a large language model is used to generate a new target script based on the new script generation prompt information to replace the original target script corresponding to the to-be-adjusted set.

[0143] Optionally, the target script generation requirement includes at least one of the following: a target object, a target object category, at least one object under a target theme, and at least one object category under a target theme, wherein the object includes a commodity or at least one multimedia material.

[0144] It should be noted that other corresponding descriptions of the functions of the script generation device based on the large language model provided in the embodiments of the present application can refer to the corresponding descriptions in the methods of FIGS. 1 to 2, which will not be described here.

[0145] The embodiments of the present application also provide a computer device, which can be a personal computer, a server, a network device, etc. The computer device includes a bus, a processor, a memory, and a communication interface, and can also include an input / output interface and a display device. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store location information. The network interface of the computer device is used to communicate with external terminals through network connection. The computer program is executed by the processor to implement the steps in the method embodiments.

[0146] Those skilled in the art can understand that the structure of the computer device described above is only part of the structure related to the scheme of the present application, and does not constitute a limitation on the computer device to which the scheme of the present application is applied. A specific computer device can include more or fewer components, or combine certain components, or have a different arrangement of components.

[0147] In an embodiment, a computer readable storage medium is provided, which can be non-volatile or volatile, and has stored thereon a computer program, which, when executed by a processor, implements the steps in the above method embodiments.

[0148] In an embodiment, a computer program product is provided, comprising a computer program, which, when executed by a processor, implements the steps in the above method embodiments.

[0149] 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 the present application are all information and data authorized by the user or authorized by all parties.

[0150] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer readable storage medium, and when the computer program is executed, the processes of the above-mentioned embodiments of the methods can be included. Any reference to memory, database or other medium used in the embodiments provided in the present application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (Read-Only Memory, ROM), magnetic tape, floppy disk, flash memory, optical storage, high-density embedded non-volatile memory, resistive memory (ReRAM), magnetoresistive random access memory (Magnetoresistive Random Access Memory, MRAM), ferroelectric memory (Ferroelectric Random Access Memory, FRAM), phase change memory (Phase Change Memory, PCM), graphene memory, etc. Volatile memory can include random access memory (Random Access Memory, RAM) or external cache memory, etc. As an illustration but not limitation, the RAM can be in various forms, such as static random access memory (Static Random Access Memory, SRAM) or dynamic random access memory (Dynamic Random Access Memory, DRAM), etc. The database involved in the embodiments provided in the present application can include at least one of a relational database and a non-relational database. The non-relational database can include a distributed database based on a block chain, etc., without being limited thereto. The processor involved in the embodiments provided in the present application can be a general-purpose processor, a graphics processor, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, etc., without being limited thereto.

[0151] Any combination of the technical features of the above embodiments can be made. In order to make the description simple, all possible combinations of the technical features in the above embodiments are not described, however, as long as the combination of the technical features does not exist contradictory, it should be considered as the scope of the present application.

[0152] The above embodiments only express several implementation manners of the present application, and the description is more specific and detailed, but it should not be understood as a limitation on the scope of the patent of the present application. It should be pointed out that for ordinary skilled in the art, without departing from the concept of the present application, a number of modifications and improvements can be made, which are within the scope of protection of the present application. Therefore, the protection scope of the present application should be subject to the appended claims.

Claims

1. A copywriting generation method based on a large language model, wherein, The method includes: Based on the target copy generation requirements, at least one data sample corresponding to at least one historical user is obtained. Each data sample includes various user characteristic data, user browsing data, and user interaction data of the corresponding historical user. The browsed object in the user browsing data matches the target copy generation requirements, and the user interaction data indicates whether an interaction is generated with the browsed object. Positive samples are constructed using data samples representing user interaction data that generated the interaction, and negative samples are constructed using the remaining data samples. A decision tree is trained using the positive and negative samples, and a user feature set is constructed based on each branch of the decision tree, wherein each user feature set includes at least one user feature type label; For any set of user features, construct copy generation prompts based on the set of user features and the target copy generation requirements, and generate the target copy based on the copy generation prompts using a large language model.

2. The method according to claim 1, wherein, Before constructing positive samples using data samples representing user interaction data that generated the interaction, and constructing negative samples using the remaining data samples, the method includes: For continuous features in user feature data, feature bucketing is performed to obtain multiple feature buckets corresponding to each continuous feature; The discrete features in the user feature data and the feature bucket are used as training features. One-hot encoding is performed on each user feature data based on the training features to obtain the user coding features corresponding to each user feature data. Accordingly, positive samples are constructed using data samples representing user interaction data that generated the interaction, and negative samples are constructed using the remaining data samples, including: Positive samples are constructed using the user coding features corresponding to the data samples that represent the user interaction data that generated the interaction, and negative samples are constructed using the user coding features corresponding to the remaining data samples.

3. The method according to claim 1, wherein, A decision tree is trained using the positive and negative samples, and a user feature set is constructed based on each branch of the decision tree, including: Based on a preset feature importance condition, a decision tree is trained using the positive and negative samples. The decision tree includes multiple branches, each branch comprising a leaf node and at least one decision section. Each decision node represents a user feature type that meets the preset feature importance condition, and each leaf node is labeled with its classification label. The target branch in each branch is determined based on the classification label of the leaf node, and a user feature set is constructed based on the user feature type represented by each decision node in each target branch.

4. The method according to claim 3, wherein, Each leaf node is also labeled with the number of samples split from the root node to that leaf node; based on the user feature types represented by each decision node in each target branch, user feature sets are constructed respectively, including: Determine the proportion of the number of samples corresponding to the leaf nodes in the target branch to the total number of data samples, and filter the target branch based on the proportion; By utilizing the user feature types represented by each decision node in each of the selected target branches, user feature sets are constructed respectively.

5. The method according to any one of claims 1 to 4, wherein, After generating the target text based on the text-generating prompt information using a large language model, the method further includes: When a target text trigger signal based on a target user is detected, the target display text is determined from the target texts corresponding to each user feature set based on the target user characteristics corresponding to the target user. The target display text is delivered to the target user.

6. The method according to claim 5, wherein, Upon detecting a target text-triggered display signal based on a target user, the target display text is determined from the target texts corresponding to each user feature set based on the target user characteristics corresponding to the target user, including: When a display signal is triggered based on the target text for the target user, the target user feature set that matches the target user feature is queried from the user feature set based on the target user feature corresponding to the target user, and the target text associated with the target user feature set is determined as the target display text. If none of the user feature sets match the target user features, then the standard copy corresponding to the target copy generation requirement will be used as the target display copy.

7. The method according to claim 5, wherein, After delivering the target display text to the target user, the method further includes: Acquire user behavior data after each type of target copy reaches users, and determine the copywriting effect of each type of target copy based on the user behavior data corresponding to each type of target copy; Based on the described copywriting effect, identify the copywriting elements in the target copywriting that do not meet the preset copywriting effect and adjust or replace them.

8. The method according to claim 7, wherein, Adjustments and replacements to the text to be adjusted include: Based on the user feature set corresponding to the text to be adjusted, the target text generation requirements, and the adjustment requirements corresponding to the text to be adjusted, a text adjustment prompt message is constructed, and an adjusted text is generated based on the text adjustment prompt message using a large language model to replace the text to be adjusted.

9. The method according to claim 7, wherein, Adjustments and replacements to the text to be adjusted include: The user feature set corresponding to the text to be adjusted is taken as the set to be adjusted. The user behavior data corresponding to the text to be adjusted is obtained as the retraining sample. A new decision tree is trained using the retraining sample. A new user feature set is constructed based on each branch of the new decision tree to replace the set to be adjusted. For any new set of user features, new copywriting generation prompts are constructed based on the new set of user features and the target copywriting generation requirements. New target copywriting is then generated based on the new copywriting generation prompts using a large language model to replace the original target copywriting corresponding to the set to be adjusted.

10. A copywriting generation device based on a large language model, wherein, The device includes: The data processing module is used to obtain at least one data sample corresponding to at least one historical user according to the target copywriting generation requirements. Each data sample includes multiple user characteristic data, user browsing data and user interaction data of the corresponding historical user. The browsed object in the user browsing data matches the target copywriting generation requirements, and the user interaction data indicates whether an interaction is generated with the browsed object. The model training module is used to construct positive samples using data samples representing user interaction data that generated the interaction, and to construct negative samples using the remaining data samples; and to train a decision tree using the positive samples and the negative samples. The model parsing module is used to construct user feature sets based on each branch of the decision tree, wherein each user feature set includes at least one user feature type label; The copywriting generation module is used to construct copywriting generation prompts based on the user feature set and the target copywriting generation requirements for any user feature set, and generate target copywriting based on the copywriting generation prompts using a large language model.

11. A storage medium having a computer program stored thereon, wherein, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 9.

12. A computer device, comprising a storage medium, a processor, and a computer program stored on the storage medium and executable on the processor, wherein, When the processor executes the computer program, it implements the method of any one of claims 1 to 9.

Images