Intelligent copywriting creation method and system based on AIGC

Abstract

The invention discloses an intelligent copywriting creation method and system based on AIGC, and relates to the technical field of calculation models, the system comprises a data acquisition unit, a model establishment unit, a copywriting generation unit, a copywriting evaluation unit and a copywriting optimization unit; the AIGC model is obtained through recurrent neural network training, the primary push copywriting is automatically generated according to the push purpose, the push display mode and the preference keywords, and the comparison keywords are extracted according to the primary push copywriting. The document content with the highest reading rate is retrieved from historical click records and historical reading records of the user to serve as a comparison document, feature extraction is conducted on the comparison document and the primary push document to conduct document evaluation, and the AIGC model is corrected according to the document evaluation result to obtain an optimized AIGC model; and the pushing purpose, the pushing display mode and the preference keywords are input into the optimized AIGC model again, the optimized pushing copywriting is obtained, a personalized copywriting generation process is provided, and the copywriting generation accuracy is improved.

Description

Technical Field

[0001] The present invention relates to the field of computing model technology, and in particular to an AIGC-based intelligent copywriting method and system. Background Art

[0002] With the popularization of mobile smart terminal devices and the rapid development of mobile applications, users often browse news, videos and other information through smart terminals. In order to improve the user experience, each application often generates and sends push copy to push hot information to users for easy browsing.

[0003] Excellent push copy can improve users' favorability towards mobile applications and increase click-through rates. However, there are still huge challenges in how to customize push copy for different users and increase users' click-through rates. Therefore, it is necessary to provide solutions with better user experience and higher recall rate. In view of the above-mentioned technical defects, a solution is now proposed. Summary of the invention

[0004] The purpose of the present invention is to perform copy evaluation by extracting features from comparison copy and primary push copy, and to modify the AIGC model according to the copy evaluation results to obtain an optimized AIGC model, thereby obtaining an optimized push copy, providing a personalized copy generation process, and improving the accuracy of copy generation.

[0005] In order to achieve the above object, the present invention adopts the following technical solution: an intelligent copywriting method based on AIGC, comprising the following steps: Step 1: Obtain a push task, which includes a push target, a push purpose, and a push display method, and simultaneously obtain the push target's personal information, which includes historical click records, historical reading records, and preferred keywords; Step 2: Obtain multiple sets of historical copywriting data, train the AIGC model through a recurrent neural network, and automatically generate primary push copywriting based on the push purpose, push display method, and preferred keywords; Step 3: extracting comparison keywords from the primary push copy, and searching for comparison keywords from the user's historical click records and historical reading records to obtain a comparison database, calculating the reading rate of each copy in the comparison database, and obtaining the copy with the highest reading rate as the comparison copy; Step 4: Extract features from the comparison copy and the primary push copy respectively to obtain feature vectors F1 and F2, and obtain the copy evaluation results by calculating vector similarity; Step 5: Modify the AIGC model according to the copy evaluation results to obtain the optimized AIGC model, and input the push purpose, push display method and preferred keywords into the optimized AIGC model again to obtain the optimized push copy.

[0006] Furthermore, the specific process of obtaining the AIGC model through recurrent neural network training is as follows: S101, obtaining multiple groups of historical copy generation data, wherein the historical copy generation data includes historical push purposes, historical push display modes, historical push copy and preferred keywords corresponding to the copy, splitting the historical copy generation data into words, assigning a unique identifier to each word, and integrating them to obtain training samples; S102, dividing the training samples into a training set and a test set in a ratio of 8:2, and performing data completion, anomaly cleaning and normalization on all data; S103, construct a neural network model containing an LSTM layer, set an embedding layer, and use a Dropout layer to randomly disconnect a certain proportion of network connections to avoid overfitting; S104, connect the LSTM layer and specify a certain proportion of Dropout; S105, connect a fully connected layer and a softmax layer, and output the primary push copy; S106, set the optimizer to adam stochastic gradient descent, set the loss function to multi-classification cross entropy function, set the measurement indicator to accuracy, and iterate multiple times until the model converges; S107. When the number of training times reaches the threshold and the loss function is greater than the historical minimum, the optimal model is selected as the AIGC model, and the training set data is used for training. The mean square error between the model result and the actual result is calculated using the test set as the model evaluation.

[0007] Furthermore, the specific process of obtaining the copywriting content with the highest reading rate as the comparison copywriting is as follows: S201, obtaining the generated primary push copy, pre-processing the text data of the primary push copy, performing word segmentation processing based on the domain dictionary, and calculating the comprehensive score of each word; S202, sorting the comprehensive scores of each word from large to small, and selecting words with scores greater than a preset score threshold as comparison keywords; S203, searching for comparison keywords from the user's historical click records and historical reading records, taking historical texts with keywords as candidate databases, performing secondary search in the candidate databases, and taking historical texts with keyword frequencies greater than a preset frequency threshold as comparison databases; S204, obtaining the historical click count Mi and historical reading time Ti corresponding to each copy content in the comparison database, and calculating the reading rate Ui of the copy content according to the following formula: , where α and β are preset weight coefficients, and the reading rate Ui is used to reflect the reading interest of the corresponding copy content of the push target. The higher the reading rate Ui; S205. Sort the reading rates of each copy content from high to low, and use the copy content ranked first as the comparison copy.

[0008] Furthermore, the specific process of calculating the comprehensive score of each word is as follows: Get the word length e, word position d and word frequency-inverse document frequency value IDF of each word after word segmentation, and calculate the comprehensive score K of each word according to the following formula: , where the word length e is calculated according to the preset proportional coefficient, different weights are assigned to words at different positions in the text content, the word position d is calculated according to the position of the word in the text content to obtain the position score, and the comprehensive score K reflects the key proportion of each word in the text data. The larger the comprehensive score K, the higher the key of the word in the text data, and vice versa, the smaller the comprehensive score K, the lower the key of the word in the text data.

[0009] Furthermore, the specific process of obtaining the copywriting evaluation results is as follows: S301, using TfidfVectorizer in sklearn, extract features from the comparison copy and the primary push copy in turn to obtain feature vectors F1 and F2. , ; S302, calculate the vector similarity S (F1, F2) according to the following formula: ; S303: obtaining a preset similarity threshold. If the similarity is greater than or equal to the similarity threshold, the copy evaluation result is obtained that the similarity between the primary push copy and the comparison copy meets the push requirement. If the similarity is less than the similarity threshold, the copy evaluation result is that the similarity between the primary push copy and the comparison copy does not meet the push requirements, and the primary push copy and the comparison copy are used as training samples.

[0010] Furthermore, the specific process of obtaining the optimized AIGC model is as follows: S401, obtaining primary push copy and comparison copy and corresponding feature vectors F1 and F2, and performing feature numerical processing on each element in the feature vectors F1 and F2; S402, using LabelEncoder to digitize the character data, mapping the label to the interval [0, n_classes-1] for encoding, and assigning a unique identifier to each label; S403, after the features are quantified, the data is standardized, and the data is uniformly linearly mapped to the interval [0, 1]. Specifically, the StandardScale algorithm is used to standardize the data, and the optimized training samples are integrated; S404, adjusting the parameters of the softmax layer and the fully connected layer, and outputting the optimized push copy; S405. Use the optimized training samples to perform multiple iterations again until the model converges. When the number of training times reaches a threshold and the loss function is greater than the historical minimum, the optimal model is selected as the optimized AIGC model.

[0011] The present invention also provides an intelligent copywriting system based on AIGC, including a data acquisition unit, a model building unit, a copywriting generation unit, a copywriting evaluation unit and a copywriting optimization unit; The data acquisition unit is used to acquire a push task, which includes a push target, a push purpose, and a push display mode, and acquires personal information of the push target, which includes historical click records, historical reading records, and preferred keywords; The model building unit is used to obtain multiple sets of historical copy generation data, obtain the AIGC model through recurrent neural network training, and automatically generate primary push copy according to the push purpose, push display method and preferred keywords; The copy generation unit is used to extract comparison keywords based on the primary push copy, and retrieve the comparison keywords from the user's historical click records and historical reading records to obtain a comparison database, calculate the reading rate of each copy content in the comparison database, and obtain the copy content with the highest reading rate as the comparison copy; The copy evaluation unit is used to extract features from the comparison copy and the primary push copy, respectively, to obtain feature vectors F1 and F2, and to obtain the copy evaluation result by calculating the vector similarity; The copy optimization unit is used to modify the AIGC model according to the copy evaluation results to obtain an optimized AIGC model, and then input the push purpose, push display method and preferred keywords into the optimized AIGC model to obtain optimized push copy.

[0012] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are: The AIGC-based intelligent copywriting method and system obtain an AIGC model through recurrent neural network training, automatically generate primary push copywriting according to push purpose, push display mode and preferred keywords, extract comparison keywords according to the primary push copywriting, retrieve the copywriting content with the highest reading rate from the user's historical click records and historical reading records as comparison copywriting, perform feature extraction on the comparison copywriting and the primary push copywriting for copywriting evaluation, correct the AIGC model according to the copywriting evaluation result to obtain an optimized AIGC model, input the push purpose, push display mode and preferred keywords into the optimized AIGC model again to obtain optimized push copywriting, provide a personalized copywriting generation process, and improve the accuracy of copywriting generation. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 The overall method flow diagram of the present invention is shown; Figure 2 The overall structural diagram of the present invention is shown. DETAILED DESCRIPTION

[0014] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0015] Embodiment 1:

[0016] like Figure 1 As shown, the intelligent copywriting creation method based on AIGC includes the following steps: Step 1: Get the push task, which includes the push target, push purpose and push display method, and get the personal information of the push target, which includes historical click records, historical reading records and preferred keywords; Step 2: Obtain multiple sets of historical copywriting data, train the AIGC model through a recurrent neural network, and automatically generate primary push copywriting based on the push purpose, push display method, and preferred keywords; The specific process of obtaining the AIGC model through recurrent neural network training is as follows: S101, obtaining multiple groups of historical copy generation data, the historical copy generation data including historical push purposes, historical push display methods, historical push copy and preferred keywords corresponding to the copy, splitting the historical copy generation data into words, assigning a unique identifier to each word, and integrating them to obtain training samples; S102, dividing the training samples into a training set and a test set in a ratio of 8:2, and performing data completion, anomaly cleaning and normalization on all data; S103, construct a neural network model containing an LSTM layer, set an embedding layer, and use a Dropout layer to randomly disconnect a certain proportion of network connections to avoid overfitting; S104, connect the LSTM layer and specify a certain proportion of Dropout; S105, connect a fully connected layer and a softmax layer, and output the primary push copy; S106, set the optimizer to adam stochastic gradient descent, set the loss function to multi-classification cross entropy function, set the measurement indicator to accuracy, and iterate multiple times until the model converges; S107. When the number of training times reaches the threshold and the loss function is greater than the historical minimum, the optimal model is selected as the AIGC model, and the training set data is used for training. The mean square error between the model result and the actual result is calculated using the test set as the model evaluation.

[0017] Step 3: extracting comparison keywords from the primary push copy, and searching for comparison keywords from the user's historical click records and historical reading records to obtain a comparison database, calculating the reading rate of each copy in the comparison database, and obtaining the copy with the highest reading rate as the comparison copy; The specific process of obtaining the copy with the highest reading rate as the comparison copy is as follows: S201, obtaining the generated primary push copy, pre-processing the text data of the primary push copy, performing word segmentation processing based on the domain dictionary, and calculating the comprehensive score of each word; The specific process of calculating the comprehensive score of each word is as follows: Get the word length e, word position d and word frequency-inverse document frequency value IDF of each word after word segmentation, and calculate the comprehensive score K of each word according to the following formula: , where the word length e is calculated according to the preset proportional coefficient, different weights are assigned to words at different positions in the text content, the word position d is calculated according to the position of the word in the text content to obtain the position score, and the comprehensive score K reflects the key proportion of each word in the text data. The larger the comprehensive score K is, the higher the key of the word in the text data is, and vice versa, the smaller the comprehensive score K is, the lower the key of the word in the text data is; S202, sorting the comprehensive scores of each word from large to small, and selecting words with scores greater than a preset score threshold as comparison keywords; S203, searching for comparison keywords from the user's historical click records and historical reading records, taking historical texts with keywords as candidate databases, performing secondary search in the candidate databases, and taking historical texts with keyword frequencies greater than a preset frequency threshold as comparison databases; S204, obtaining the historical click count Mi and historical reading time Ti corresponding to each copy content in the comparison database, and calculating the reading rate Ui of the copy content according to the following formula: , where α and β are preset weight coefficients, and the reading rate Ui is used to reflect the reading interest of the corresponding copy content of the push target. The higher the reading rate Ui; S205. Sort the reading rates of each copy content from high to low, and use the copy content ranked first as the comparison copy.

[0018] Step 4: Extract features from the comparison copy and the primary push copy respectively to obtain feature vectors F1 and F2, and obtain the copy evaluation results by calculating vector similarity; The specific process of obtaining the copywriting evaluation results is as follows: S301, using TfidfVectorizer in sklearn, extract features from the comparison copy and the primary push copy in turn to obtain feature vectors F1 and F2, , ; S302, calculate the vector similarity S (F1, F2) according to the following formula: ; S303: obtaining a preset similarity threshold. If the similarity is greater than or equal to the similarity threshold, the copy evaluation result is obtained that the similarity between the primary push copy and the comparison copy meets the push requirement. If the similarity is less than the similarity threshold, the copy evaluation result is that the similarity between the primary push copy and the comparison copy does not meet the push requirements, and the primary push copy and the comparison copy are used as training samples.

[0019] Step 5: Modify the AIGC model according to the copy evaluation results to obtain the optimized AIGC model, and input the push purpose, push display method and preferred keywords into the optimized AIGC model again to obtain the optimized push copy.

[0020] The specific process of obtaining the optimized AIGC model is as follows: S401, obtaining primary push copy and comparison copy and corresponding feature vectors F1 and F2, and performing feature numerical processing on each element in the feature vectors F1 and F2; S402, using LabelEncoder to digitize the character data, mapping the label to the interval [0, n_classes-1] for encoding, and assigning a unique identifier to each label; S403, after the features are quantified, the data is standardized, and the data is uniformly linearly mapped to the interval [0, 1]. Specifically, the StandardScale algorithm is used to standardize the data, and the optimized training samples are integrated; S404, adjusting the parameters of the softmax layer and the fully connected layer, and outputting the optimized push copy; S405. Use the optimized training samples to perform multiple iterations again until the model converges. When the number of training times reaches a threshold and the loss function is greater than the historical minimum, the optimal model is selected as the optimized AIGC model.

[0021] The present invention obtains an AIGC model through recurrent neural network training, automatically generates a primary push copy according to a push purpose, a push display mode and preferred keywords, extracts comparison keywords according to the primary push copy, retrieves the copy content with the highest reading rate from the user's historical click records and historical reading records as the comparison copy, performs feature extraction on the comparison copy and the primary push copy to perform copy evaluation, corrects the AIGC model according to the copy evaluation result to obtain an optimized AIGC model, inputs the push purpose, the push display mode and the preferred keywords into the optimized AIGC model again to obtain an optimized push copy, provides a personalized copy generation process, and improves the accuracy of copy generation.

[0022] Embodiment 2:

[0023] like Figure 2 As shown, the intelligent copywriting system based on AIGC includes a data acquisition unit, a model building unit, a copywriting generation unit, a copywriting evaluation unit and a copywriting optimization unit; The data acquisition unit is used to acquire the push task, which includes the push target, push purpose and push display mode, and acquire the personal information of the push target, which includes the historical click record, historical reading record and preferred keywords; The model building unit is used to obtain multiple sets of historical copy generation data, obtain the AIGC model through recurrent neural network training, and automatically generate primary push copy according to the push purpose, push display method and preferred keywords; The copy generation unit is used to extract comparison keywords based on the primary push copy, and retrieve the comparison keywords from the user's historical click records and historical reading records to obtain a comparison database, calculate the reading rate of each copy content in the comparison database, and obtain the copy content with the highest reading rate as the comparison copy; The copy evaluation unit is used to extract features from the comparison copy and the primary push copy, respectively, to obtain feature vectors F1 and F2, and to obtain the copy evaluation result by calculating the vector similarity; The copy optimization unit is used to modify the AIGC model according to the copy evaluation results to obtain an optimized AIGC model, and then input the push purpose, push display method and preferred keywords into the optimized AIGC model to obtain optimized push copy.

[0024] The threshold is set to facilitate comparison. The size of the threshold depends on the amount of sample data and the number of bases set by technicians in this field for each group of sample data; as long as it does not affect the proportional relationship between the parameter and the quantized value.

[0025] The above formulas are all dimensionless and numerical calculations. The formula is a formula obtained by collecting a large amount of data and performing software simulation to obtain the most recent real situation. The preset parameters in the formula are set by technicians in this field according to actual conditions. In the two embodiments provided in the present application, it should be understood that the disclosed method and system can be implemented in other ways; for example, the division of the modules is only a logical function division, and there may be other division methods in actual implementation, such as multiple modules or components can be combined or integrated into another system, or some features can be ignored or not executed; another point, the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces; The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. The intelligent copywriting creation method based on AIGC is characterized by: The following steps are involved: Step 1: Obtain a push task, which includes a push target, a push purpose, and a push display method, and simultaneously obtain the push target's personal information, which includes historical click records, historical reading records, and preferred keywords; Step 2: Obtain multiple sets of historical copywriting data, train the AIGC model through a recurrent neural network, and automatically generate primary push copywriting based on the push purpose, push display method, and preferred keywords; Step 3: extracting comparison keywords from the primary push copy, and searching for comparison keywords from the user's historical click records and historical reading records to obtain a comparison database, calculating the reading rate of each copy in the comparison database, and obtaining the copy with the highest reading rate as the comparison copy; Step 4: Extract features from the comparison copy and the primary push copy respectively to obtain feature vectors F1 and F2, and obtain the copy evaluation results by calculating vector similarity; Step 5: Modify the AIGC model according to the copy evaluation results to obtain the optimized AIGC model, and input the push purpose, push display method and preferred keywords into the optimized AIGC model again to obtain the optimized push copy.

2. The AIGC-based intelligent copywriting method according to claim 1, characterized in that: The specific process of obtaining the AIGC model through recurrent neural network training is as follows: S101, obtaining multiple groups of historical copy generation data, wherein the historical copy generation data includes historical push purposes, historical push display modes, historical push copy and preferred keywords corresponding to the copy, splitting the historical copy generation data into words, assigning a unique identifier to each word, and integrating them to obtain training samples; S102, dividing the training samples into a training set and a test set in a ratio of 8:2, and performing data completion, anomaly cleaning and normalization on all data; S103, construct a neural network model containing an LSTM layer, set an embedding layer, and use a Dropout layer to randomly disconnect a certain proportion of network connections to avoid overfitting; S104, connect the LSTM layer and specify a certain proportion of Dropout; S105, connect a fully connected layer and a softmax layer, and output the primary push copy; S106, set the optimizer to adam stochastic gradient descent, set the loss function to multi-classification cross entropy function, set the measurement indicator to accuracy, and iterate multiple times until the model converges; S107. When the number of training times reaches the threshold and the loss function is greater than the historical minimum, the optimal model is selected as the AIGC model, and the training set data is used for training. The mean square error between the model result and the actual result is calculated using the test set as the model evaluation.

3. The AIGC-based intelligent copywriting method according to claim 1, characterized in that: The specific process of obtaining the copy with the highest reading rate as the comparison copy is as follows: S201, obtaining the generated primary push copy, pre-processing the text data of the primary push copy, performing word segmentation processing based on the domain dictionary, and calculating the comprehensive score of each word; S202, sorting the comprehensive scores of each word from large to small, and selecting words with scores greater than a preset score threshold as comparison keywords; S203, searching for comparison keywords from the user's historical click records and historical reading records, taking historical texts with keywords as candidate databases, performing secondary search in the candidate databases, and taking historical texts with keyword frequencies greater than a preset frequency threshold as comparison databases; S204, obtaining the historical click count Mi and historical reading time Ti corresponding to each copy content in the comparison database, and calculating the reading rate Ui of the copy content according to the following formula: , where α and β are preset weight coefficients, and the reading rate Ui is used to reflect the reading interest of the corresponding copy content of the push target. The higher the reading rate Ui; S205. Sort the reading rates of each copy content from high to low, and use the copy content ranked first as the comparison copy.

4. The AIGC-based intelligent copywriting method according to claim 1, characterized in that: The specific process of calculating the comprehensive score of each word is as follows: Get the word length e, word position d and word frequency-inverse document frequency value IDF of each word after word segmentation, and calculate the comprehensive score K of each word according to the following formula: , where the word length e is calculated according to a preset proportional coefficient, and different weights are given to words in different positions in the text content.

5. The AIGC-based intelligent copywriting method according to claim 1, characterized in that: The specific process of obtaining the copywriting evaluation results is as follows: S301, using TfidfVectorizer in sklearn, extract features from the comparison copy and the primary push copy in turn to obtain feature vectors F1 and F2. , ; S302, calculate the vector similarity S (F1, F2) according to the following formula: ; S303: obtaining a preset similarity threshold. If the similarity is greater than or equal to the similarity threshold, the copy evaluation result is obtained that the similarity between the primary push copy and the comparison copy meets the push requirement. If the similarity is less than the similarity threshold, the copy evaluation result is that the similarity between the primary push copy and the comparison copy does not meet the push requirements, and the primary push copy and the comparison copy are used as training samples.

6. The AIGC-based intelligent copywriting method according to claim 1, characterized in that: The specific process of obtaining the optimized AIGC model is as follows: S401, obtaining primary push copy and comparison copy and corresponding feature vectors F1 and F2, and performing feature numerical processing on each element in the feature vectors F1 and F2; S402, using LabelEncoder to digitize the character data, mapping the label to the interval [0, n_classes-1] for encoding, and assigning a unique identifier to each label; S403, after the features are quantified, the data is standardized, and the data is uniformly linearly mapped to the interval [0, 1]. Specifically, the StandardScale algorithm is used to standardize the data, and the optimized training samples are integrated; S404, adjusting the parameters of the softmax layer and the fully connected layer, and outputting the optimized push copy; S405. Use the optimized training samples to perform multiple iterations again until the model converges. When the number of training times reaches a threshold and the loss function is greater than the historical minimum, the optimal model is selected as the optimized AIGC model.

7. The AIGC-based intelligent copywriting system is characterized by: It includes data acquisition unit, model building unit, copywriting generation unit, copywriting evaluation unit and copywriting optimization unit; The data acquisition unit is used to acquire a push task, which includes a push target, a push purpose, and a push display mode, and acquires personal information of the push target, which includes historical click records, historical reading records, and preferred keywords; The model building unit is used to obtain multiple sets of historical copy generation data, obtain the AIGC model through recurrent neural network training, and automatically generate primary push copy according to the push purpose, push display method and preferred keywords; The copy generation unit is used to extract comparison keywords based on the primary push copy, and retrieve the comparison keywords from the user's historical click records and historical reading records to obtain a comparison database, calculate the reading rate of each copy content in the comparison database, and obtain the copy content with the highest reading rate as the comparison copy; The copy evaluation unit is used to extract features from the comparison copy and the primary push copy, respectively, to obtain feature vectors F1 and F2, and to obtain the copy evaluation result by calculating the vector similarity; The copy optimization unit is used to modify the AIGC model according to the copy evaluation results to obtain an optimized AIGC model, and then input the push purpose, push display method and preferred keywords into the optimized AIGC model to obtain optimized push copy.

Images