JP7245478B1 - CONTENT GENERATION METHOD, CONTENT PROVISION METHOD, AND THEREOF PROGRAM - Google Patents

Abstract

An object of the present invention is to have AI automatically create content for distribution with a high conversion rate (CVR) and provide it to specified target consumers. A basic content generation step of generating basic content by a trained GPT model that has performed machine learning on task learning data; and a trained BERT model that has performed machine learning on the task learning data. An elaboration step of elaborating the base content to randomly hide words in the content and provide alternative candidates, and feed back to the elaboration step, leaving only the candidates that increase the CVR in a neural network that predicts CVR. causing a computer to execute a feedback step; and a content-for-distribution generation step of repeating the elaboration step and the feedback step to generate, as content-for-distribution, content whose CVR has converged. [Selection drawing] Fig. 1

Description

The present invention relates to a content generation method, a content provision method and a program thereof, and more particularly to a content generation method, a content provision method and a program thereof that meet the purpose of marketing.

In today's information explosion era, the amount of time people can pay attention to things in a day has reached its limit. ” has economic value. As such, as the economy moves from a consumption economy to an information economy, consumer attention is transforming into a valuable resource, and attention is becoming something like currency. On the other hand, when drafting advertisements such as texts and push notifications related to products and services, attention is paid to itself rather than the superiority or inferiority of the information provided to consumers.

It should be noted that it is not easy to continue to grasp the interest of consumers in the vast amount of information, especially the creation of advertising proposals such as traditional sentences and push notifications, which are mostly proposed and carried out by humans. Therefore, it is difficult to respond quickly to a dizzying information environment. On the other hand, Patent Literature 1 discloses a technique in which a deep neural network model component trains according to past product description information and example content of product description information to generate content corresponding to product description information with AI. there is As a result, it has become possible to quickly generate distribution contents for target products, so new contents can be quickly provided even in rapidly changing information environments. However, the inventors of the present application felt that simply providing new product-related content quickly with AI would only quickly notify consumers of new product sales. Specifically, even if training is performed according to past product description information and example content of the product description information, content that focuses on the product itself tends to be generated. In other words, although a large amount of content for distribution can be generated, there is a high possibility that content that is not of interest to consumers will be generated.

Further, Patent Document 2 discloses a technique of uploading an image for distribution automatically created by AI with keywords and images to an advertisement bulletin board point, and selecting the image for distribution based on the evaluation of the image at the advertisement bulletin board point. . However, the evaluation at the advertisement bulletin board point is sparse, so the efficiency of selection is low, and the value of the result is not stable. In addition, since it is to converge the number of images for distribution, it may be possible to select items that are relatively interesting to consumers from a large amount of content, but it takes time and effort to repeat the same work over and over again. However, since the quality of the content remains the same overall, the visitor's CVR cannot be substantially increased.

International Application Publication No. WO2019/133545 Taiwan Patent Publication No. TW201918960A

The problem to be solved by the present invention is a content generation method, a content provision method, and a method for automatically creating distribution content with a high conversion rate (CVR) and natural sentences by AI and providing it to specified target consumers. provide that program.

A content generation method according to one embodiment includes an acquisition step of acquiring appeal information, and a basic content generation step of generating basic content based on the appeal information by a trained GPT model that has performed machine learning on task learning data. and an elaboration step of elaborating the basic content so as to randomly hide words in the basic content and give alternative candidates by a trained BERT model that has performed machine learning on task learning data; a feedback step of predicting a conversion rate of the text elaborated by the elaboration step with a trained NN (neural network) model trained, leaving only the candidates that increase the conversion rate, and providing feedback to the elaboration step; The step and the feedback step are repeated to cause a computer to execute a content-for-distribution generation step of generating content for which the increase in conversion rate has converged as content for distribution.

A method for providing content according to one embodiment includes steps of predicting a type of content for distribution by a trained neural network model for predicting the type of content, and associating the content for distribution with the predicted type; Predicting the type of content that is likely to convert to the target target, which is a task, by a trained NN model that predicts the type of content that is likely to be converted, and associating the target target with the predicted type; Correspondingly providing the distribution content to the target according to the associated type.

A program according to one embodiment obtains appeal information, generates basic content based on the appeal information by a trained GPT model that has performed machine learning on task learning data, and performs machine learning on task learning data. The basic content is refined by randomly hiding words in the basic content by a trained BERT model and providing alternative candidates, and a trained NN (neural network) model trained by task prediction learning data is used. Predicting the conversion rate of the sentence refined by the BERT model, leaving only the candidates that increase the conversion rate of the sentence, feeding back the BERT model refinement, and repeating the BERT model refinement and the NN model feedback step. Then, the computer executes a process of generating the content for which the increase in the conversion rate has converged as the content for distribution.

A program according to one embodiment predicts the type of content for distribution generated by the aforementioned program by a trained NN model for predicting the type of content, and associates the content for distribution with the predicted type. , using a trained NN model that predicts the type of content that is likely to convert to the target, predicts the type of content that is likely to convert to the target, which is the task, and predicts the type predicted as the target A computer is caused to execute a process of associating and correspondingly providing the distribution content to the target according to the associated type.

According to the above-described embodiment of the present invention, AI automatically creates content for distribution of natural sentences with a high conversion rate (CVR), and at the timing determined by AI for the target that is easy to CV determined by AI can be output with

4 is a flow chart of an exemplary flow of a content generation method according to an embodiment of the present invention; 4 is an image of a GPT model generating basic content, which is used in the content generation method according to the embodiment of the present invention; 4 is an image of a BERT model used in a content generation method according to an embodiment of the present invention to generate refined content; 1 is an illustration of the general architecture of a NN model used in a content generation method according to an embodiment of the invention; FIG. FIG. 4 is an explanatory diagram showing an example of past scenario data included in learning data according to the embodiment of this invention; FIG. 4 is an explanatory diagram showing an example of appeal information provided by a client, which is included in learning data according to the embodiment of this invention; FIG. 4 is an explanatory diagram showing an example of sentences obtained by searching the Internet for appealing information included in learning data according to the embodiment of the present invention; FIG. 4 is a diagram for explaining an embodiment of task prediction learning data used in the NN model of the embodiment of the present invention; FIG. FIG. 4 is a diagram for explaining the flow of all services after receiving a request from a client using the content generation method of the embodiment of the present invention; 4 is a flow chart of an exemplary flow of a method for providing content according to an embodiment of the present invention; Fig. 2 is an illustration of the general architecture of a fine-tuned type predictive NN model and a fine-tuned user predictive NN model used in the content provision method according to an embodiment of the present invention; To illustrate embodiments of learning data including sentence-type task prediction learning data and user-type task prediction learning data used in fine-tuned type prediction NN models and fine-tuned user prediction NN models according to embodiments of the present invention; is a diagram. FIG. 2 is a diagram for explaining the flow of all services for providing distribution content according to an embodiment of the present invention; 1 is a structural diagram of a hardware configuration that can implement a content generation method and a content provision method according to an embodiment of the present invention; FIG.

Embodiments relating to the "content generating method, content providing method, and program therefor" disclosed by the present invention will be described below with specific examples. Those skilled in the art can understand the advantages and effects of the present invention from the disclosure of this specification. The invention may be practiced or applied with other different embodiments. Each detail in this specification can also be changed based on various applications without departing from the spirit of the invention. Also, the drawings of the present invention are for simple and schematic illustration and do not show actual dimensions. In the following embodiments, technical matters related to the present invention will be further described, but the disclosed contents do not limit the present invention.

Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs. Technical terms used herein are intended to include the plural unless otherwise specified.

Preferred embodiments for carrying out the present invention will now be described in detail with reference to the accompanying drawings. It should be noted that such an embodiment is merely an example and does not limit the present invention.

According to one aspect of the present invention, a content generation method is provided. An exemplary flow of a content generation method in an embodiment of the present invention will now be described with reference to FIG.

FIG. 1 is a flowchart of an exemplary flow of a content generation method according to an embodiment of the invention. As shown in FIG. 1, an exemplary content generation method flow 100 may include the following steps.

Step S101: fine-tuning the pre-trained GPT model and the pre-trained BERT model according to the task learning data;
Step S103: Fine-tuning a pre-trained NN (neural network) model for predicting conversion rate according to the task prediction learning data;
Step S105: Giving the fine-tuned GPT model the transcripts extracted from the sentences of the task learning data to generate sentences;
Step S107: The sentence generated in step S105 is input to the fine-tuned BERT model to extract the distributed representation, and the sentence having a high cosine similarity with the distributed representation of the sentence of the task learning data before extraction in step S105 is used as the basis. Extract as content (basic sentences);
Step S109: Input the basic sentences extracted in step S107 into the fine-tuned BERT model, hide words at random, and list alternative candidates to generate revised sentences;
Step S111: Predict the conversion rate of the elaboration sentences generated in step S109 by the fine-tuned NN model, leave only candidates for elaboration contents (elaboration sentences) whose conversion rate rises , and use those candidates as new basic sentences. and feedback to the fine-tuned BERT model;
Step S113: In step S111, it is determined whether the increase in the conversion rate of the revised text has converged. If it is determined that the increase in the conversion rate has converged, the process proceeds to step S115. When judging, it returns to step S109.
Step S115: Output the revised text as distribution contents (distribution text).

In this flow 100, each detail, such as details of task learning data and task prediction learning data, will be described later with examples. When using a model, it is obvious for those skilled in the art to convert the format according to the machine learning model used, such as converting the task learning data to a sequence of vectors before using it. It may be performed according to the method, and the related description is omitted in this specification.

In recent years, research in the field of natural language processing of deep learning has been actively conducted, and among them, the BERT model and the GPT model are widely used. The GPT model performs fine tuning of parameters of an existing model and additional learning, called fine tuning, on a pre-learned model, and makes highly accurate predictions without reconstructing a new model. Especially for the generation of long sentences, the GPT model can produce sentences that are natural enough to be judged as written by humans. However, since the sentences generated by the GPT model simply arrange words based on past information and create grammatically appropriate sentences, when generating long sentences, words with the same meaning may be repeated or conclusions may not be reached. It may produce contradictory sentences. On the other hand, the BERT model enables interactive pre-training using Transformer, so that highly accurate and contextually understood expressions can be obtained. That is, in the content generation method according to the present application, the BERT model, which is good at rewriting (filling in blanks) of sentences, is used in addition to the GPT model, which is good at writing the continuation of sentences, in order to automatically create sentences for distribution. As a result, natural sentences are automatically created as if they were created by a human being.

Then, in order to make AI create not only natural sentences like those created by humans, but also sentences for distribution that achieve marketing-related goals such as increasing CVR, reducing CPA, or increasing LTV, the present invention can be used. Such a content generation method further utilizes a neural network model for predicting conversion rates. Specifically, the BERT distributed representation of the sentence to be predicted is input to a pretrained NN model that predicts the conversion rate, and the conversion rate (value between 0 and 1) is output. If the output conversion rate increases, we feed back the candidates used for that sentence to the BERT model.

In that way, the BERT model can be made to produce high conversion rate sentences until the conversion rate increase converges. As long as the NN model used here can predict the conversion rate, any existing neural network such as a forward propagation neural network (FFNN), a convolutional neural network (CNN), a recurrent neural network (RNN), etc. You can use it.

Subsequently, based on FIGS. 2 to 4, more specific examples of fine-tuning and mechanics of the GPT model, BERT model and NN model applied to the exemplary method shown in FIG. 1 will be described.

First, refer to FIG. FIG. 2 is an image of generating basic sentences in the GPT model, which is used in the content generation method shown in FIG. In this embodiment, the first GPT model used is ckiplab/gpt2-base-chinese uploaded to Huggingface Transformers as a GPT pre-learning model. The invention is not limited to this example, as it may vary accordingly. In this embodiment, the pre-trained GPT model may be fine-tuned before executing the content generation method. For example, fine-tuning may use a GPT model that was trained last month. That is, in the content generation method of the present invention, the pre-trained GPT model and the fine-tuned GPT model are divided depending on whether or not fine-tuning is performed using new task learning data when the content generation method of this time is started. decide.

In one embodiment, the task learning data used for fine-tuning the GPT model is past scenario data, appeal information provided by the client, and sentences obtained by searching the Internet for appeal information, as will be described later. included. A sentence in the present application is a series of sentences including text, pictograms or images to express a cohesive idea. Pictograms or images are converted to a format that the GPT model can handle. For example, pictograms may be converted to characters according to a pre-provided list. Images may be converted to text by existing image analysis methods. In addition, appropriately selected data from past task learning data is treated as test data.

After the task learning data is input to the pretrained GPT model, fine-tuning is performed. The number of fine-tuning epochs is not specified, but 5 epochs is preferred. Other training parameters are default values for TrainingArguments in Huggingface transformers. After confirming that the loss of both the learning data and the inspection data has decreased, they are saved as a fine-tuned GPT model. In an embodiment according to the present invention, if the loss of inspection data has turned from a decrease to an increase, the model immediately before the increase is saved as the fine-tuned GPT model.

In this way, when the fine-tuned GPT model is provided with the writing of task learning data sentences, it is possible to automatically create desirable basic sentences.

After understanding the fine-tuning and mechanics of the GPT model shown in FIG. 2, a practical example of BERT model training is subsequently described based on FIG.

First, refer to FIG. FIG. 3 is an image of how the BERT model used in the content generation method shown in FIG. 1 generates the revised content. In this embodiment, hfl/chinese-bert-wwm-ext uploaded to Huggingface Transformers was used as the BERT pre-learning model to be used first, but the pre-learning model to be used depends on the language to be generated. It can be modified and the invention is not limited to this example. Note that in this embodiment, the pre-trained BERT model may be fine-tuned before the content generation method is executed. For example, fine-tuning may use the BERT model that was fine-tuned last month. That is, in the content generation method of the present invention, the pre-trained BERT model and the fine-tuned BERT model are divided depending on whether or not fine-tuning is performed using new task learning data at the time of starting the content generation method of this time. decide.

In this embodiment, fine tuning of the BERT model is the same as fine tuning of the GPT model, including details of task learning data and inspection data.

After the task learning data is input to the BERT model, fine-tuning is performed. The number of fine-tuning epochs is not specified, but 5 epochs is preferred. Other training parameters are default values for TrainingArguments in Huggingface transformers. After confirming that the loss of both the training data and the inspection data is decreasing, save it as a fine-tuned BERT model. Save the model as a fine-tuned BERT model.

In this way, when the fine-tuned BERT model is given the sentence of the basic sentence generated by the GPT model, in the basic sentence, words are randomly hidden and the fine-tuned BERT selects alternative candidates. and outputs a distributed representation of the revised text.

After understanding the fine-tuning and mechanics of the BERT model shown in FIG. 3, an example of constructing a fine-tuned NN model for predicting conversion rates based on FIG. 4 will be described.

Please refer to FIG. FIG. 4 is an explanatory diagram of the general architecture of the NN model used in the content generation method shown in FIG. As shown in FIG. 4, the NN model used in the exemplary method shown in FIG. 1 includes NN40. It basically includes an input layer 401 , a hidden layer 402 and an output layer 403 . The input layer 401 consists of N neurons for inputting an N-dimensional vector, the hidden layer 402 consists of N/2 neurons, and the output layer 403 consists of 1 neuron. Here, in the present invention, N is a positive integer unless otherwise specified. In addition, in this embodiment, there is one hidden layer 402 and N is 768, but both the number of hidden layers and the value of N can be adjusted according to actual needs. do not have. Also, in this embodiment, the activation function used for the neurons in the hidden layer 402 is ReLU, and the activation function used for the neurons in the output layer 403 is the sigmoid function. It is not intended to be limited to this example, as it may be adjusted accordingly.

It should be noted that those skilled in the art should be able to understand the principle, structure and specific details of the NN model. An example of fine-tuning will be briefly explained. In one example, the input layer 401 of the NN 40 from the left side shown in FIG. ). In the content generation method of this embodiment, a total of 24 neural networks having the same configuration as the NN 40 are used in the NN model, but the present invention is not limited to this example.

In one preferred embodiment, scenario sentences and conversion rates are obtained from past scenario data as task prediction learning data, as will be described later. Input task prediction learning data into a pre-trained BERT model. From each layer of the BERT model (12 layers excluding the word embedding layer), the distributed representation of the beginning [CLS] and ending [SEP] of the sentence is taken out. A total of 12×2=24 distributed representations (all 768-dimensional vectors) are obtained. Each type of distributed representation is input to the 24 pre-trained NNs 40, and each output is obtained. Calculate the difference between a total of 24 outputs and the actual conversion rate, and fine-tune the neural network so that the sum of the absolute values approaches zero. When estimating the conversion rate after fine-tuning, we use the average value of a total of 24 outputs.

Next, an example of using the fine-tuned NN model for the content generation method of this embodiment will be described. The conversion rate is predicted by the fine-tuned NN model for the sentences refined by the fine-tuned BERT model, and only the candidates that increase the conversion rate are left and fed back to the fine-tuned BERT model. By iterating elaboration with the fine-tuned BERT model and feedback of conversion rate boosting candidates, it is possible to produce elaboration sentences that converge on predicted conversion rate increases. These revised texts are output as texts for distribution (contents for distribution), and the process of the content generation method according to this embodiment ends.

Please refer to FIGS. 5-7. 5-7 are diagrams for explaining embodiments of task learning data used for the BERT model and the GPT model. In this embodiment, the task learning data includes past scenario data, appealing information provided by the client, and/or sentences obtained by searching the appealing information on the Internet.

FIG. 5 is an explanatory diagram showing an example of past scenario data included in the task learning data of this embodiment. As shown in FIG. 5, past scenario data includes, for example, conversation records with target users in public or business chat rooms, and data obtained through conversations with users such as interview results. Collection, classification, and storage of such past scenario data may be performed by any of the existing data collection/processing methods as long as a specific number of sentences can be obtained with specific classification in a specific period.

FIG. 6 is an explanatory diagram showing an example of the appeal information provided by the client, which is included in the task learning data of this embodiment. A client here means a person who requests a content creation service by the content creation method of the present invention. For example, advertisers, administrative agencies, research institutes, etc. can be mentioned. The content of the appeal information differs depending on the client, but at least includes sentences related to the ideas the client wants to convey and statistical information related to the target user. For example, FIG. 6 shows an example of appeal information assuming that the client is an advertiser. In FIG. 6, examples of sentences related to the idea that the client wants to convey include product introduction, side reference, advertising appeal with good results, and the like. Details of targets such as the elderly are included. In the case of request, acquisition of appeal information may be conducted by taking a questionnaire to the client so as to include the above-mentioned information, or by any other existing data collection/processing method.

FIG. 7 is an explanatory diagram showing an example of sentences obtained by searching the Internet for appealing information included in the task learning data of the present embodiment. In order to further enrich the expression of the sentences created by the content generation method of the present invention, an Internet search is performed using some key information obtained from the appeal information, and, for example, the sentences of the top 10 ranking sites are extracted. can be used as sentences obtained by searching the Internet for appeal information. Here, in order to further refine the sentences obtained by searching the Internet for the appeal information, the distributed representation of the pre-trained BERT model is used. For example, it is possible to select those whose cosine similarity is equal to or higher than a specific threshold. In this embodiment, the specific threshold is preferably 0.9. It should be noted that sentences obtained by searching the Internet for appeal information may be replaced with standard sentences related to appeal information, but this is not preferable because it may deviate from the latest sentence expressions.

Next, please refer to FIG. FIG. 8 is a diagram for explaining an embodiment of task prediction learning data used in the NN model. In the present embodiment, the task prediction learning data is data configured in relation to past scenario sentences and conversion rates . For example, in the table shown in FIG. 8, column B indicates the scenario text, column E the number of people who read the scenario text, and column G the number of people who purchased (CV) the scenario text. Column A displays the identification number of the sentence and the purchase screen, and Column C displays the identification number of the sentence read before the sentence (or the purchase screen). For example, 593 people read the text in column 2, and 38 of them made a purchase (as shown in column 3), so the conversion rate is 38/593 = 6.41%. Become. If the scenario sentence consists of multiple sentences, train using a string of those sentences. As long as a plurality of past scenario sentences, conversion rates , and their corresponding relationships can be acquired, any existing information processing method can be used as a method or data expression format for acquiring such data. .

The flow and details of the content generation method according to the embodiment of the present invention have been described above with reference to FIGS. 1 to 8 . Next, based on this example, an example of a service that provides the content generation method according to the present invention will be described with reference to FIG.

FIG. 9 is a diagram for explaining the flow of all services after receiving a request from a client.
First, take a questionnaire prepared for the client and obtain appeal information.
Then, search for the appealing information on the Internet and acquire the sentences of the top 10 sites displayed, along with the appealing information and past scenario data within a certain time period (for example, up to 2 months before), task learning Summarize as data.
Fine-tuning the pre-trained GPT model and the pre-trained BERT model with the task learning data.
Fine-tuning a pre-trained NN model to predict conversion rates with the latest task prediction learning data.
The fine-tuned GPT model is provided with the writing out of task learning data sentences (5 to 10 characters) to automatically generate sentences.
Of the sentences created by the fine-tuned GPT model, sentences close to the sentences of the appeal information are selected as basic sentences. This selection uses the variance representation of the fine-tuned BERT model and chooses the one with the highest cosine similarity.
In the basic sentence, the words are randomly hidden, and the fine-tuned BERT model performs elaboration to select alternative candidates. Re-execute refinement by feeding back to the existing BERT model.
Repeat the above elaboration and feedback until the expected conversion rate increase converges. The document so refined is generated as a distribution document, and the process ends.
Through the above process, a plurality of distribution documents can be generated.

In this way, from the client's request, AI can automatically generate natural distribution sentences that increase the conversion rate. Can be more closely related to readers (i.e. consumers, target users), ultimately reaching marketing-related goals such as increasing CVR, reducing CPA or increasing LTV be able to. In addition, the distribution text in the present invention includes, for example, pop-up advertisements, materials for chatbot conversations, push notification messages, etc., but the present invention is not limited to these examples.

Next, with reference to FIG. 10, an example of a content providing method for providing distribution content generated by the content generating method according to the present invention to target users will be described. FIG. 10 is a flow chart of an exemplary flow of a method for providing content according to an embodiment of the present invention. As shown in FIG. 10, an exemplary flow 1000 of a method for providing content may include the following steps.

Step S1001: Perform fine-tuning of the pre-trained type prediction NN model for predicting the type of content (sentence) by the sentence type task prediction learning data;
Step S1003: Perform fine-tuning of the pre-trained user prediction NN model for predicting the type of sentences that are likely to be CVed to the target user by the user type task prediction learning data;
Step S1005: Input the distribution text to be predicted into the fine-tuned type prediction NN model, predict the type of the distribution text, and associate the type with the distribution text;
Step S1007: Input the target user data into the fine-tuned user prediction NN model, predict the type of distribution text that is likely to be CVed to the target user, and associate the type with the target user;
Step S1009: providing each distribution text to each target user according to its associated type;

In this way, according to the content providing method of the present invention, it is possible to provide a target user with a type of sentence that is easy to CV.

Next, based on FIG. 11, an actual example of configuration of a type prediction NN model for predicting the type of content and a user prediction NN model for predicting the type of sentences that target users are likely to CV will be described.

Please refer to FIG. FIG. 11 is an illustration of the general architecture of the type predictive NN model and the user predictive NN model used in the content provision method shown in FIG. As shown in FIG. 11, the structures of the type predictive NN model and the user predictive NN model used in the exemplary method shown in FIG. 10 also include a type predictive NN 1100. Type prediction NN 1100 basically includes input layer 1101 , hidden layer 1102 and output layer 1103 . The input layer 1101 consists of N neurons for inputting an N-dimensional vector, the hidden layer 1102 consists of N/2 neurons, and the output layer 1103 consists of 4 neurons. Here, in the present invention, N is a positive integer unless otherwise specified. In the present embodiment, the hidden layer 1102 is one layer, N is 768, and the number of neurons in the output layer is four. It is not intended to be limited to this example as it may be adjusted according to needs. Also, in this embodiment, the activation function used for the neurons of the hidden layer 1102 is ReLU, and the activation function used for the neurons of the output layer 1103 is the sigmoid function. It is not intended to be limited to this example, as it may be adjusted accordingly.

It should be noted that those skilled in the art should be able to understand the principle, structure and specific details of the NN model. An example of fine-tuning a model and a pre-trained user-predictive NN model is briefly described. In one example, first, in the fine-tuning of the pre-trained type prediction NN model, the input layer 1101 of the type prediction NN 1100 from the left side shown in FIG. (only the distributed representation of the deepest layer [CLS] is used), and from the output layer 1103 composed of the number of types (here, 4) of neurons in the sentence type task prediction learning data, the predicted sentences are each is trained to output probabilities (values between 0 and 1) that are of the type Similarly, in the fine tuning of the pretrained user prediction NN model, user type task prediction learning data is input to the input layer 1101 of the type prediction NN 1100 from the left side shown in FIG. An output layer 1103 consisting of a number (here 4) of neurons is trained to output the probability that the target user will convert in each type of scenario ( conversion rate, value between 0 and 1).

Next, an example of using the fine-tuned type predictive NN model and the fine-tuned user predictive NN model will be described. Specifically, in one preferred embodiment, the handout sentences we wish to predict, as described below, are input to a pre-trained BERT model, and the distributed representation of the BERT model (the deepest sentence head [CLS] distributed representation only). The distributed representation of the BERT model is then input into the fine-tuned type prediction NN model, and the type with the highest probability obtained is associated with the handout text. In one preferred embodiment, target user data, as described below, is input into a fine-tuned user prediction NN model, and the type with the highest probability obtained is mapped to the target user. When distributing texts are provided to target users, they are provided according to the type associated with each.

Next, refer to FIGS. 8 and 12. FIG. Figures 8 and 12 are provided to illustrate embodiments of learning data including sentence-type task prediction learning data and user-type task prediction learning data used in fine-tuned type prediction NN models and fine-tuned user prediction NN models. is a diagram. In this embodiment, the learning data includes past scenario sentences, sentence types, and target user data.

In the table shown in FIG. 8, sentence type task prediction learning data and user type task prediction learning data are registered. For example, in column B of the table, scenario text is displayed, and in column F, the type of text is displayed as text type task prediction learning data. In this embodiment, the types are price superiority (scenario emphasizing that the product is cheap), effect superiority (scenario emphasizing that the product is highly effective), and influencer superiority (scenario emphasizing that the product is used by celebrities). Scenario to emphasize), conversion rate superiority (scenario that encourages customers to repurchase products that they have already purchased). On the other hand, in the table shown in FIG. 12, user type task prediction learning data is registered. For example, column A is a list of options, column B is the number of users who chose it, column C is the number of users who converted (CV), and column D and onwards are the number of users who converted in each scenario. In addition, read the type of each scenario sentence. By combining the above data, it is possible to calculate the probability that users who chose each option converted with what type of sentence. This becomes user type task prediction learning data. The collection, classification, and storage of these data may be performed by any of the existing data collection/processing techniques.

FIG. 13 is a diagram for explaining the flow of all services for providing content for distribution.
Fine-tuning of the pretrained type prediction NN model and the pretrained user prediction NN model is performed using the latest sentence type task prediction learning data and user type task prediction learning data as learning data.
In order to select targets to send handouts, we used a fine-tuned type predictive NN model and a fine-tuned user prediction NN model to determine which handouts sent from the choices each target user made in the past had the highest conversion rate. Predict whether it will be easy and choose.
Selected by the above process, the text for distribution that is most likely to convert for each target user is sent to each target user.

It should be noted that after sending, the data of the scenarios read by each user, the choices made in the scenarios, and which scenarios led to conversions remain on the server and may be used as the next learning data. In addition, according to the content providing method according to the present embodiment, it is also possible to provide distribution texts to achieve other purposes according to the client's wishes. For example, in another embodiment, if the goal of improving CVR is to increase customer lifetime value (LTV), for example, in addition to the content provision method in the above embodiment, for example, access to the target user's landing page Various goals can be served, such as further use of NN models for predicting when to send out handouts (eg, push notifications) with high rates.

An information processing apparatus is provided as another aspect of the present invention. The processor of the information processing device performs operations and/or processes of, for example, the content generation method and the content provision method described with reference to FIGS. However, detailed description thereof is omitted here. The NN model used by the processor may also be obtained by the exemplary process for optimizing the architecture of the model described with reference to FIGS. 4 and 11 above.

FIG. 14 is a structural diagram of a hardware configuration (general-purpose machine) 1400 that can implement the content generation method, content provision method, and information processing apparatus according to the embodiment of the present invention.

General purpose machine 1400 may be, for example, a computer system. It should be noted that general purpose machine 1400 is exemplary only and does not limit the scope of application or functionality of the methods and apparatus according to the present invention. Also, general-purpose machine 1400 does not rely on any modules, assemblies, etc., or combinations thereof in the methods and apparatus described above.

In FIG. 14, a central processing unit (CPU) 1401 performs various processes based on programs stored in a ROM 1402 or programs loaded from a storage unit 1408 to a RAM 1403 . The RAM 1403 can also store data necessary for the CPU 1401 to perform various processes according to needs. The CPU 1401 , ROM 1402 and RAM 1403 are interconnected via a bus 1404 . Input/output interface 1405 is also connected to bus 1404 .

The input/output interface 1405 is further connected with the following components: an input unit 1406 including a keyboard, an output unit including a display such as a liquid crystal display (LCD) and a speaker. 1407, a storage unit 1408 including a hard disk, etc., and a communication unit 1409 including a network interface card such as a LAN card, a modem, and the like. A communication unit 1409 performs communication processing via a network such as the Internet or a LAN, for example. Drives 1410 may be connected to input/output interfaces 1405 as desired. A removable medium 1411 , such as a semiconductor memory, can be set in the drive 1410 as necessary to install a computer program read therefrom into the storage unit 1408 .

Additionally, the present invention further provides a program product comprising machine-readable instruction code. Such instruction code, when read and executed by a machine, is capable of performing the methods in the embodiments of the invention described above. Correspondingly, various storage media such as magnetic discs, optical discs, magneto-optical discs, and semiconductor memories carrying such program products are also included in the present invention.

The above storage medium may include, for example, a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory device, etc., but is not limited to these.

Also, each operation (process) in the above-described method can be realized in the form of a computer-executable program stored in various machine-readable storage media.

Although the preferred embodiment of the present invention has been described above, the present invention is not limited to this embodiment, and all modifications to the present invention fall within the technical scope of the present invention as long as they do not depart from the gist of the present invention.

40 fine-tuned NN
401 Input layer 402 Hidden layer 403 Output layer 1100 Type prediction NN
1101 Input layer 1102 Hidden layer 1103 Output layer 1400 General-purpose machine 1401 Central processing unit 1402 ROM
1403 RAM
1404 bus 1405 input/output interface 1406 input unit 1407 output unit 1408 storage unit 1409 communication unit 1410 drive 1411 removable medium

Claims (10)

an acquisition step of acquiring solicitation information;
a basic content generation step of generating basic content based on the appeal information by a trained GPT model that has performed machine learning with task learning data;
an elaboration step of elaborating the base content to randomly hide words in the base content and suggest alternative candidates by a trained BERT model that has performed machine learning on task learning data;
Predicting the conversion rate of the sentences elaborated by the elaboration step with a trained NN (neural network) model trained by task prediction learning data, leaving only the elaborated sentences that increase the conversion rate, and using them as new basis a feedback step for treating content as feedback to the elaboration step;
a content-for-distribution generation step of repeating the elaboration step and the feedback step to generate content for which the increase in conversion rate has converged as content for distribution;
A content generation method characterized by causing a computer to execute 2. The task learning data includes past scenario data, the appeal information, or high-ranking content obtained by searching a network for the appeal information, which includes the content of a conversation with a target target by a chatbot. content generation method described in . Distributed representation of task prediction learning data by a pre-trained BERT model is used for training the NN model, and the task prediction learning data is data configured in relation to the past scenario data and conversion rates. 2. The content generation method according to 2. 3. The content generation method according to claim 2, wherein said task learning data includes content whose cosine similarity with said past scenario data exceeds a predetermined threshold among data extracted from the Internet. The task learning data includes data extracted from the Internet or data prepared in advance in a database,
5. The content generation method according to claim 4, wherein the content whose cosine similarity with said past scenario data exceeds a certain threshold is extracted by measuring the cosine similarity of the content using a trained BERT model. 6. The content generation method according to any one of claims 1 to 5, wherein said task prediction learning data includes at least content containing an idea that an appealer wants to convey and statistical data of a target that the appealer wants to convey. The content generation method according to any one of claims 1 to 5, wherein said content for distribution is a message used for conversation by a chatbot or a message for push notification. The type of content for distribution generated by the content generation method according to any one of claims 1 to 7 is predicted by a trained NN model for predicting the type of content, and the type of content for distribution is predicted as the content for distribution. mapping a type;
A trained NN model that predicts the type of content that is likely to convert to the target is used to predict the type of content that is likely to be converted to the target, which is a task, and the predicted type is associated with the target. a step;
Correspondingly providing the content for distribution to the target according to the associated type;
A content providing method characterized by causing a computer to execute to obtain solicitation information,
Generate basic content based on the appeal information by a trained GPT model that has performed machine learning with task learning data,
elaborating the base content by randomly hiding words in the base content and suggesting alternative candidates by a trained BERT model that has performed machine learning on task learning data;
A trained NN (neural network) model trained with task prediction learning data predicts the conversion rate of the sentences refined by the BERT model, leaving only the refined sentences that increase the conversion rate of the sentences, and leaving them Regarded as new basic content, feed back to the elaboration of the BERT model,
Repeating the refinement of the BERT model and the feedback step of the NN model to generate content for which the increase in conversion rate has converged as content for distribution;
A content generation program characterized by causing a computer to execute processing. Predicting the type of content for distribution generated by the program according to claim 9 by a trained NN model for predicting the type of content, and associating the content for distribution with the predicted type,
Using a trained NN model that predicts the type of content that is likely to convert to the target, predicts the type of content that is likely to convert to the target, which is the task, and responds to the predicted type of target Attached to,
Correspondingly providing the distribution content to the target according to the associated type;
A content providing program characterized by causing a computer to execute processing.

Images