A method for constructing promotional copy training corpus and related devices

Abstract

The present application discloses a method and related device for constructing a training corpus of promotional copywriting. The method classifies the relevant corpus by identifying the types of promotional products and the types of copywriting styles in the relevant training corpus of the promotional copywriting, determines a plurality of style tags including the category combination of the promotional product types and the copywriting styles, and then obtains the corresponding professional knowledge and the corresponding training corpus for each style tag in the plurality of style tags according to the product type in the style tag, and then generates prompt words according to the style tag, the corresponding professional knowledge and the corresponding training corpus, and then inputs the prompt words into a first generation model to generate a training corpus sample of the promotional copywriting corresponding to the target style tag, so that the promotional copywriting obtained based on the training sample has higher accuracy, appeal and professionalism for different promotion scenarios, can achieve better promotion effect, and effectively improves the quality of the training corpus and the efficiency of constructing the training corpus.

Description

Technical Field

[0001] The present application relates to the field of machine learning technology, and in particular to a method for constructing a promotional copy training corpus and related devices. Background Art

[0002] Promotional copy is the key link between promoted products and consumers. The professionalism, compliance and appeal of promotional copy directly affect the conversion effect of promoted products.

[0003] By constructing the promotion copy training corpus, it can be input into the big model to generate the promotion copy. In the related art, the construction method of the promotion copy training corpus is mainly that the promotion experts write sample copy and annotate the professional knowledge.

[0004] However, the cost of manually writing training corpora is high, making the construction of training corpora inefficient. Summary of the invention

[0005] In view of this, the present application provides a method and related devices for constructing a large model training corpus of marketing copy. By introducing professional knowledge and relevant copy design prompts for different promotional products and copy styles, a large number of training samples are obtained for different promotion needs, thereby effectively improving the efficiency of constructing the promotion copy training corpus.

[0006] To solve the above problems, the technical solutions provided by this application are as follows:

[0007] On the one hand, the present application provides a method for constructing a promotional copy training corpus, the method comprising:

[0008] Obtain the first training corpus of promotional copy;

[0009] Extracting professional features and copywriting style features of the first training corpus, wherein the professional features are used to identify the promotion product type of the first training corpus, and the copywriting style features are used to identify the copywriting style of the first training corpus;

[0010] Classifying the first training corpus according to the professional features and the copywriting style features to determine a plurality of style tags, wherein the style tags are used to indicate a category combination of a promotional product type and a copywriting style;

[0011] Generate an indicator word according to a target style tag among the multiple style tags, a training corpus corresponding to the target style tag, and professional knowledge corresponding to the target style tag;

[0012] The indicator word is input into a first generation model to generate a training sample corresponding to the target style label, and the training sample is used to train a promotional copy recognition model, and the promotional copy recognition model is used to recognize the promotional product type and copy style of the input copy to be recognized.

[0013] In a possible implementation, the professional knowledge includes professional terminology, promotional product information, and compliance content.

[0014] In a possible implementation, after inputting the indicator word into the first generation model to generate a training sample corresponding to the target style label, the method further includes:

[0015] Determine reference copy that meets the promotion needs corresponding to the target style tag;

[0016] Inputting the prompt word into a second generation model to generate text corresponding to the target style label, wherein the scale of the second generation model is smaller than that of the first generation model;

[0017] Evaluate the document according to the reference document to obtain an evaluation result;

[0018] The prompt word is optimized according to the evaluation result.

[0019] In a possible implementation, the evaluation result includes a professional evaluation result and a promotion effect evaluation result.

[0020] In a possible implementation, the classifying the first training corpus according to the professional features and the copywriting style features to determine a plurality of style tags includes:

[0021] Based on the attention mechanism model, the professional features and the copywriting style features are fused to obtain fused features;

[0022] The relevant corpus is classified according to the fusion features to determine the multiple style tags.

[0023] In a possible implementation, the optimizing the prompt word according to the evaluation result includes:

[0024] The evaluation results of the two dimensions of professionalism and promotion effect are analyzed. If the evaluation result of the target dimension does not meet the expected value, the content corresponding to the target dimension in the prompt word is adjusted. The target dimension is any one or both of the two dimensions of professionalism and promotion effect.

[0025] On the other hand, the present application provides a device for constructing a promotional copy training corpus, characterized in that the device includes an acquisition unit, an extraction unit, a determination unit, and a generation unit:

[0026] The acquisition unit is used to acquire a first training corpus of the promotional copy;

[0027] The extraction unit is used to extract professional features and copywriting style features of the first training corpus, wherein the professional features are used to identify the promotion product type of the first training corpus, and the copywriting style features are used to identify the copywriting style of the first training corpus;

[0028] The determining unit is used to classify the first training corpus according to the professional characteristics and the copywriting style characteristics, and determine a plurality of style labels, wherein the style labels are used to indicate a category combination of a promotional product type and a copywriting style;

[0029] The generating unit is used to generate an indicator word according to a target style tag among the multiple style tags, a training corpus corresponding to the target style tag, and professional knowledge corresponding to the target style tag;

[0030] The generation unit is also used to input the indicator word into the first generation model to generate training corpus corresponding to the target style label. The training sample is used to train the promotional copy recognition model, and the promotional copy recognition model is used to identify the promotion product type and copy style of the input copy to be recognized.

[0031] In a possible implementation, the professional knowledge includes professional terminology, promotional product information, and compliance content.

[0032] In a possible implementation manner, the generating unit is further configured to:

[0033] Determine reference copy that meets the promotion needs corresponding to the target style tag;

[0034] Inputting the prompt word into a second generation model to generate text corresponding to the target style label, wherein the scale of the second generation model is smaller than that of the first generation model;

[0035] Evaluate the document according to the reference document to obtain an evaluation result;

[0036] The prompt word is optimized according to the evaluation result.

[0037] In a possible implementation, the evaluation result includes a professional evaluation result and a promotion effect evaluation result.

[0038] In a possible implementation manner, the determining unit is specifically configured to:

[0039] Based on the attention mechanism model, the professional features and the copywriting style features are fused to obtain fused features;

[0040] The relevant corpus is classified according to the fusion features to determine the multiple style tags.

[0041] In a possible implementation manner, the generating unit is further configured to:

[0042] The evaluation results of the two dimensions of professionalism and promotion effect are analyzed. If the evaluation result of the target dimension does not meet the expected value, the content corresponding to the target dimension in the prompt word is adjusted. The target dimension is any one or both of the two dimensions of professionalism and promotion effect.

[0043] In another aspect, the present application provides a computer device, characterized in that the computer device includes a processor and a memory:

[0044] The memory is used to store computer programs;

[0045] The processor is configured to execute the method according to the computer program.

[0046] On the other hand, the present application provides a computer-readable medium, characterized in that the computer-readable medium is used to store a computer program, and the computer program implements the method when executed by a computer device.

[0047] In yet another aspect, the present application provides a computer program product comprising a computer program, characterized in that when the computer program product is run on a computer device, the computer device is enabled to perform the method.

[0048] It can be seen that this application has the following beneficial effects:

[0049] The present application provides a method for constructing a promotional copy training corpus. First, relevant training expectations of the promotional copy are obtained, and professional features and copy style features corresponding to different promotional product types are extracted from the relevant corpus. Then, the relevant corpus is classified according to the extracted features to determine a plurality of style labels including category combinations of promotional product types and copy styles. Then, each of the plurality of style labels is processed separately. For a target style label among the plurality of style labels, corresponding professional knowledge and corresponding training corpus are obtained according to the product type in the target style label. Then, prompt words are generated according to the target style label, the corresponding professional knowledge and the corresponding training corpus. Finally, the prompt words are input into a first generation model to generate a training sample of the promotional copy corresponding to the target style label.

[0050] Therefore, by classifying the style labels of relevant training corpus, different promotional product types and promotional speech styles can be distinguished, and then the product type, copywriting style, professional knowledge related to the product type and the relevant training corpus corresponding to the style labels after classification are used as guidance to generate prompt words, so that the generated prompt words can match the promotion needs of the corresponding product type and speech style, and the professional knowledge of the promotion product can improve the professionalism of the prompt words, thereby effectively improving the guiding role of the prompt words, so that the accuracy, appeal and professionalism of the training corpus promotion products can be improved, and better promotion effects can be achieved. In addition, through the generation of large models, a large number of corresponding promotional copywriting training samples can be generated according to specific promotion needs, thereby effectively improving the efficiency of constructing promotional copywriting training corpus compared with the method of manually constructing training corpus. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0052] Figure 1 A flowchart of a method for constructing a promotional copy training corpus provided in an embodiment of the present application;

[0053] Figure 2 A schematic diagram of constructing a promotional copy training corpus provided in an embodiment of the present application;

[0054] Figure 3 A schematic diagram of a flow chart of a device for constructing a promotional copy training corpus provided in an embodiment of the present application. DETAILED DESCRIPTION

[0055] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0056] As described in the background technology, the existing method of constructing promotional copy training corpus is mainly for promotion experts to write sample copy and annotate professional knowledge. This method of manually writing and annotating training corpus consumes a lot of manpower and time costs, making the construction efficiency of training corpus low. The present application provides a method for constructing promotional copy training corpus. By classifying the relevant training corpus of the promotion product, it is possible to distinguish promotion scenarios including different product types and speech styles, and then use the product type, speech style, product type-related professional knowledge and related training corpus as guidance to generate prompt words, so that the generated prompt words can match the corresponding product type and speech style promotion needs, improve the accuracy and appeal of the promotion, and the professional knowledge of the promotion product can improve the professionalism of the prompt words, thereby effectively improving the guiding role of the prompt words, so that the generated training corpus can achieve better promotion effects, and in this way, a large number of corresponding promotional copy training corpora can be generated for specific promotion scenarios, thereby effectively improving the efficiency of constructing promotional copy training corpora compared with the method of manually constructing training corpora.

[0057] The method provided in this application is applied to the field of machine learning and can be specifically applied to product promotion, which is specifically illustrated by the following embodiments.

[0058] See also Figure 1 As shown, it is a flowchart of a method for constructing a promotional copy training corpus provided in an embodiment of the present application. In this embodiment, it can be executed by a computer device as an example.

[0059] S101: Obtain the first training corpus of the promotional copy.

[0060] The first training corpus refers to text information related to the promoted product. By inputting this text information into the big model, it can help the big model analyze the relevant information of the promoted product and generate the corresponding promotional copy for the promoted product.

[0061] The computer device can retrieve relevant text information on the Internet based on keywords related to the promoted products, or integrate information such as slogans, product introductions, promotional copy, etc. set by the promoter to obtain a large amount of corpus information to form the first training corpus.

[0062] As an example, for the promotion of insurance products, the first training corpus may include text information such as product introductions, promotional slogans, and coverage introductions of life insurance or critical illness insurance. The following is an example of a critical illness insurance product introduction in the first training corpus:

[0063] “Dear young friends:

[0064] With the money you spend on a cup of milk tea every month, you can add a layer of protection to your healthy life! Our "Healthy Life" critical illness insurance covers 100 major diseases, with a maximum insured amount of up to 1 million yuan. Once unfortunately diagnosed, you can get compensation, so that you have sufficient financial support during the most difficult times of your life. If you are insured now, you can also get a critical illness hospitalization allowance, up to 300 yuan per day! "

[0065] S102: Extracting professional features and copywriting style features of the first training corpus.

[0066] Professional features refer to features related to the promoted product, and are used to identify the type of promoted product in the first training corpus. For example, a computer device can extract features of related information such as the type of promoted product, core terms, and professional terms from the first training corpus through a pre-built coding module to help identify the type of promoted product. Copywriting style features are used to identify the copywriting style of corpus information. For example, a computer device can extract features of scenario descriptions and core selling points of products in the first training corpus through a pre-built coding module to help identify the copywriting style.

[0067] As an example, taking insurance promotion products as an example, the professional features and copywriting style features extracted from a corpus in the first training corpus are as follows:

[0068] Professional characteristics = {

[0069] 'Insurance term':['Critical illness insurance', 'Major disease', 'Insurance amount', 'Claim', 'Hospitalization allowance'],

[0070] 'Product Details':{

[0071] 'Name':'XXX',

[0072] 'Type': 'Critical Illness Insurance',

[0073] 'Coverage': '100 major diseases',

[0074] 'Maximum fault reporting range':'1 million'

[0075] }},

[0076] Copywriting style characteristics

[0077] 'Emotional vocabulary': ['dear', 'unfortunate', 'ample'],

[0078] 'Target audience': 'Young people',

[0079] 'Writing style': 'Friendly'

[0080] }

[0081] S103: Classify the first training corpus according to the professional characteristics and the copywriting style characteristics to determine a plurality of style labels.

[0082] Style tags are used to indicate a combination of promotional product types and copywriting styles. A style tag includes a promotional product type and a copywriting style category. The promotional product type and copywriting style category can be pre-set according to actual needs. For example, for insurance products, promotional product types can include life insurance, medical insurance and other categories. For different target groups, copywriting style categories can include light-hearted humor, neutral and objective, etc., and the style tags include (life insurance, light-hearted humor), (life insurance, neutral and objective), (medical insurance, light-hearted humor) and (medical insurance, neutral and objective).

[0083] The computer device can classify the product type and copywriting style of each training corpus in the first training corpus through a preset network model, and determine the style label corresponding to each training corpus one by one through the type combination of product type and copywriting style corresponding to each training corpus, thereby obtaining multiple style labels based on the types of style labels contained in the first training corpus.

[0084] For example, the types of promotional products corresponding to different training corpora in the first training corpus can be determined based on the extracted professional features, and the copywriting styles corresponding to different corpora can be determined based on the extracted copywriting style features. Then, based on the product types and copywriting style categories in the first training corpus, multiple style labels can be determined by combining them respectively. The extracted professional features and copywriting style features can also be fused to jointly determine the product types and copywriting style categories of the same training corpus, thereby obtaining multiple style labels.

[0085] S104: Generate an indicator word according to a target style tag among the multiple style tags, a training corpus corresponding to the target style tag, and professional knowledge corresponding to the target style tag.

[0086] The training corpus corresponding to the target style label refers to the training corpus corresponding to the target style label in the first training corpus. After the first training corpus is classified, the training corpuses in the first training corpus have corresponding style labels respectively, and the same style label may have multiple corresponding training corpora.

[0087] The target style tag is one of multiple style tags. The professional knowledge corresponding to the target style tag refers to the professional knowledge related to the type of product promoted in the target style tag, such as professional terminology, violation statements, and coverage.

[0088] The indicator word is a text fragment or instruction input to the model, which is intended to guide the model to generate the expected output. For a target style tag among multiple style tags, the computer device can input the target style tag, multiple training corpora corresponding to the target style tag, and professional knowledge corresponding to the target style tag into a pre-designed model architecture for generating indicator words, and generate one or more indicator words, which can include multiple keywords or sentences related to the input content.

[0089] As an example, the designed model architecture can extract product information, copywriting style, target audience, promotion focus, compliance requirements and other features from the target style label, corresponding professional knowledge and corresponding training corpus, and then output keywords or sentences based on the features. The following is an example of a prompt word:

[0090] "Critical illness insurance promotion for young white-collar workers aged 25-30, XXX critical illness insurance (product name), 100 types of critical illnesses, starting at 3 yuan per day, easy conversation type."

[0091] Among them, "young white-collar workers aged 25-30" represents the target population, "XXX critical illness insurance" is the product type, "100 types of critical illnesses" and "starting from three yuan a day" represent the promotion focus, and "relaxed conversation type" represents the copywriting style.

[0092] S105: Input the indicator word into the first generation model to generate training samples corresponding to the target style label.

[0093] The training samples are used to train the promotional copy recognition model, which is used to identify the promotional product type and copy style of the input copy to be recognized. In practical applications, the training samples corresponding to the target style label may also include the corresponding training corpus obtained in step S101. The computer device inputs the indicator word into the first generation model to instruct the model to generate the training sample corresponding to the target style label.

[0094] As an example, here are the training samples generated for the prompt words in the above example:

[0095] "Working people, you who work hard for your dreams every day, have you ever thought about buying an "airbag" for your health?

[0096] XXX critical illness insurance, with the price of a cup of milk tea a day, can provide you with protection against 100 major diseases. If you are sick, you can pursue your dreams with peace of mind if you have insurance! "

[0097] Therefore, by classifying the style labels of the first training corpus, different types of promotional products and promotional speech styles can be distinguished, and prompt words can be generated guided by the product type, copywriting style, professional knowledge related to the product type, and the relevant training corpus corresponding to the style labels after classification. The generated prompt words can match the promotion needs of the corresponding product type and speech style, and the professional knowledge of the promotional products can improve the professionalism of the prompt words, thereby effectively improving the guiding role of the prompt words, so that the accuracy, appeal and professionalism of the training corpus promotion products can be improved, and better promotion effects can be achieved. In addition, a large number of corresponding promotional copywriting training samples can be generated according to specific promotion needs through the generation of large models, thereby effectively improving the efficiency of constructing promotional copywriting training corpus compared with the method of manually constructing training corpus.

[0098] In one possible approach, expertise includes specialized terminology, promotional product information, and compliance content.

[0099] Professional terminology refers to the unified name for certain specific things in the field of promotional products. Promotional product information represents the characteristics of the promotional products and is used to indicate the type of promotional products. It may include information such as product name and coverage. Compliance content refers to relevant promotional content related to the product that complies with laws and regulations.

[0100] By restricting indicator words through promoting product-related professional terms, promotion product information and related compliance content, the indicator words can contain more accurate and comprehensive information, so that the generated promotion copy training samples can achieve better results.

[0101] In a possible implementation, after the indicator word is input into the first generation model to generate a training sample corresponding to the target style label, the method further includes the following steps:

[0102] A1: Determine the reference copy that meets the promotion needs corresponding to the target style label.

[0103] Reference copy refers to the copy that can achieve better results under the promotion requirements corresponding to the target style label. The training samples generated under the promotion requirements can be evaluated through the reference copy. The reference copy can be manually set by experts according to the requirements.

[0104] A2: Input the prompt word into the second generation model to generate the text corresponding to the target style label. The scale of the second generation model is smaller than that of the first generation model.

[0105] The second generation model is used to test the promotion effect of the training samples generated according to the prompt words. The first generation model can generate a large number of training samples according to the prompt words, that is, under the same promotion demand. The computer device can select some training samples for testing, and quickly generate the copy corresponding to the target style label by inputting the prompt words into the second generation model.

[0106] A3: Evaluate the copy based on the reference copy and get the evaluation result.

[0107] After the copy corresponding to the target style tag is generated, the evaluation is performed according to the set reference copy. In order to obtain a better evaluation result, the evaluation can be performed from multiple aspects. In a possible example, the evaluation result includes a professional evaluation result and a promotion effect evaluation result.

[0108] In the actual evaluation process, two stages of evaluation were conducted. The first was the incremental pre-training stage, where the prompt word was input into the second generation model to evaluate a large number of generated promotional copywriting to optimize the model's adaptability to new training samples generated by the prompt word. Then, in the stage of fine-tuning the second generation model, the prompt word was input into the second generation model to evaluate a small number of generated promotional copywriting. The evaluation results in the incremental pre-training and fine-tuning processes can include professional evaluation results and promotion effect evaluation results.

[0109] Among them, the professional evaluation results and promotion effect evaluation results are determined by comparing the copy with the reference copy. For example, the professional evaluation results can be evaluated from three aspects: the accuracy of professional terms in the copy, the completeness of product descriptions, and the clarity of guarantee instructions, and three corresponding evaluation scores can be obtained. The promotion effect evaluation results can be evaluated from four aspects: the matching degree of the target population corresponding to the target style label, the accuracy of pain point descriptions, the clarity of value propositions, and the conversion efficiency of the product, and four corresponding evaluation results can be obtained.

[0110] As an example, the evaluation results of the fine-tuning phase are as follows:

[0111] Evaluation result = {

[0112] 'Professional evaluation results':

[0113] 'Professional term accuracy': 0.95,

[0114] 'Product description completeness': 0.92,

[0115] 'Clarity of warranty statement': 0.90

[0116] },

[0117] 'Promotion effect evaluation results':{

[0118] 'Target population matching': 0.88,

[0119] 'Pain point description accuracy': 0.85,

[0120] 'Value proposition clarity': 0.87,

[0121] 'Expected conversion rate': 0.82

[0122] }.

[0123] Therefore, through multi-faceted evaluation results, the expected effect of promoting products can be accurately controlled, so that the prompt words can be adjusted purposefully.

[0124] A4: Optimize the prompt words based on the evaluation results.

[0125] After obtaining the evaluation results, the prompt words can be optimized in a targeted manner according to actual needs. For example, when it is necessary to improve the accuracy of professional terms in the professional evaluation results, more professional terms can be added to the prompt words or certain keywords can be replaced with professional terms for more accurate expressions. When it is necessary to improve the target population matching in the promotion effect evaluation results, a specific description of the target population can be added.

[0126] Therefore, through the evaluation results of the training samples generated by the prompt words, the prompt words can be optimized in a targeted manner, so that the prompt words can be more accurate, thereby generating higher quality training samples for the promotion needs corresponding to the prompt words.

[0127] In a possible implementation, the first training corpus is classified according to the professional characteristics and the copywriting style characteristics, and the multiple style tags are determined, including:

[0128] Based on the attention mechanism model, the professional features and copywriting style features are fused to obtain fusion features;

[0129] The relevant corpus is classified according to the fusion features to determine multiple style labels.

[0130] The attention mechanism model can be a commonly used attention mechanism model or an attention mechanism model pre-built for the classification task.

[0131] By fusing professional features and copywriting style features according to the attention mechanism model, we can explore the correlation between professional features and copywriting style features, extract information related to professional features from copywriting style features, and improve the expressiveness of professional features. We can also extract information related to copywriting style features from professional features and improve the expressiveness of copywriting style features, thereby improving the performance of classification tasks and ensuring the accuracy of copywriting style and product type recognition in relevant corpus.

[0132] In a possible implementation, optimizing the prompt words according to the evaluation results includes:

[0133] Analyze the evaluation results from the two dimensions of professionalism and promotion effect. If the evaluation result of the target dimension does not meet the expected value, adjust the content corresponding to the target dimension in the prompt word.

[0134] The target dimension can be the professional dimension or the promotion effect dimension, or it can be both the professional dimension and the promotion effect dimension.

[0135] As an example, the promoter sets the expected value of the accuracy of professional terms in the professional dimension to 95%, and the expected value of the target population matching in the promotion effect dimension to 90%. If the accuracy of the professional terms in the copy obtained after the prompt word is input into the second generation model is 90%, then a more accurate professional expression will be made for the key words or sentences related to the professional terms in the prompt word. If the matching degree of the target population in the copy obtained after the prompt word is input into the second generation model is 88%, then a more detailed description will be made for the key words or sentences related to the target population in the prompt word.

[0136] Therefore, by optimizing the prompt words through the evaluation results of the two dimensions, the promotional copy generated according to the training samples can achieve the promotion effect expected by the promoter.

[0137] For a clearer description of the construction process of the promotional copy training corpus, see Figure 2 , which is a schematic diagram of constructing a promotional copy training corpus provided in an embodiment of the present application.

[0138] Taking insurance promotion products as an example, merchants need to promote three types of products: life insurance, critical illness insurance, and medical insurance, and set the copywriting style to neutral and objective, light-hearted and humorous, and rigorous and professional to face different target audiences.

[0139] First, the relevant corpus information of the three products is obtained as the first training corpus through manual writing, and then the first training corpus is classified through the copywriting style classification module to identify the promotion product types and copywriting styles corresponding to different corpora. The classification module is implemented by building a network model. The network model first extracts professional features and copywriting style features from the relevant corpus information, and then fuses the extracted features based on the feature fusion method of the attention mechanism. After that, the style classification decision is made based on the fused features to identify the type of promotion product and the copywriting style of the corpus corresponding to each corpus information in the first training corpus, thereby obtaining the style label corresponding to each corpus, that is, the type combination of promotion product type and copywriting style. Among them, professional features may include features related to promotion products such as insurance product type, product coverage, disclaimers and professional vocabulary, and copywriting style features may include value propositions, scenario descriptions, insurance claims descriptions and other features.

[0140] After obtaining the style tags corresponding to the relevant corpus information, multiple style tags included in the promotional copy can be determined, and then promotional copy training samples can be generated in the training sample generation module for different style tags. For example, for the style tags of critical illness insurance and humor, professional knowledge related to critical illness insurance can be introduced to perform knowledge injection, and compliance control can be performed by avoiding illegal expressions, and one or more prompt words can be generated through the relevant expectations in the first training expectations, and then the prompt words are input into the first generation model for sample generation to obtain training samples, and the quality of the training samples can be improved by performing quality control of the training samples based on the training effect evaluation module.

[0141] The training effect evaluation module includes incremental pre-training evaluation and supervised training evaluation. The incremental pre-training evaluation is the evaluation of the incremental pre-training of the second generation model, and the supervised training evaluation is the evaluation of the fine-tuning of the second generation model. The evaluation of incremental pre-training includes compliance evaluation and professional evaluation, and the supervised evaluation results include compliance evaluation and effect evaluation. The above evaluation results can be obtained by comparing the promotional copy generated by the second generation model with the sample copy annotated by experts.

[0142] Finally, the prompt words are optimized and fed back. After obtaining the evaluation results, the results are fed back to the training sample generation module. If the evaluation results do not meet expectations, the design of the prompt words is optimized based on the evaluation results.

[0143] The above three modules demonstrate the process of applying the promotional copy training corpus construction method to the promotion of insurance products, forming a complete technical closed loop from style label classification, sample generation to effect evaluation.

[0144] The copywriting style classification module first extracted multi-dimensional features, used a feature fusion method based on the attention mechanism to dynamically assign weights, and determined the style labels of different corpora through style classification decisions. The training sample generation module reversely generated prompt words through style labels, compliance control, and related corpus information, and then generated samples based on the prompt words. It also performed quality control through multi-faceted evaluations in the incremental pre-training phase and model fine-tuning phase of the catenary effect evaluation module, generating high-quality training samples, thereby ensuring the professionalism, compliance, and better promotion effect of the promotional copy generated based on the training samples.

[0145] In addition, in the process of executing the above method, a more comprehensive analysis is conducted on the effects that can be achieved by different modules. The effects that can be achieved by different modules are as follows:

[0146] 1. Style classification effect of the copywriting style classification module:

[0147] The style recognition accuracy rate reached 92%;

[0148] Supports automatic classification of various styles such as neutrality and objectivity, light-hearted humor, rigorous and professional;

[0149] Classification efficiency increased by 21%.

[0150] 2. Sample generation effect of training sample generation module:

[0151] Labor costs reduced by 80%;

[0152] Sample generation efficiency increased by 300%;

[0153] The professional accuracy of the generated samples reached 90%.

[0154] 3. Effect of training effect evaluation module:

[0155] Reduce assessment time by 75%;

[0156] The incremental pre-training evaluation accuracy reached 82%;

[0157] The supervised training evaluation accuracy reached 81%.

[0158] Based on the method for constructing the promotional copy training corpus provided in the above embodiments, the embodiment of the present application further provides a device for constructing the promotional copy training corpus, which is described in detail below in conjunction with the accompanying drawings.

[0159] See also Figure 3 , which is a flow chart of a device for constructing a promotional copy training corpus provided by the present application, wherein the device comprises an acquisition unit 301, an extraction unit 302, a determination unit 303 and a generation unit 304:

[0160] The acquisition unit 301 is used to acquire a first training corpus of the promotional copy;

[0161] The extraction unit 302 is used to extract professional features and copywriting style features of the first training corpus, wherein the professional features are used to identify the promotion product type of the first training corpus, and the copywriting style features are used to identify the copywriting style of the first training corpus;

[0162] The determining unit 303 is used to classify the first training corpus according to the professional characteristics and the copywriting style characteristics, and determine a plurality of style labels, wherein the style labels are used to indicate a category combination of a promotional product type and a copywriting style;

[0163] The generating unit 304 is used to generate an indicator word according to a target style tag among the multiple style tags, a training corpus corresponding to the target style tag, and professional knowledge corresponding to the target style tag;

[0164] The generating unit 304 is further used to input the indicator word into the first generating model to generate training corpus corresponding to the target style label, and the training sample is used to train the promotional copy recognition model, and the promotional copy recognition model is used to identify the promotional product type and copy style of the input copy to be recognized.

[0165] In a possible implementation, the professional knowledge includes professional terminology, promotional product information, and compliance content.

[0166] By restricting indicator words through promoting product-related professional terms, promotion product information and related compliance content, the indicator words can contain more accurate and comprehensive information, so that the generated promotion copy training samples can achieve better results.

[0167] In a possible implementation manner, the generating unit is further configured to:

[0168] Determine reference copy that meets the promotion needs corresponding to the target style tag;

[0169] Inputting the prompt word into a second generation model to generate text corresponding to the target style label, wherein the scale of the second generation model is smaller than that of the first generation model;

[0170] Evaluate the document according to the reference document to obtain an evaluation result;

[0171] The prompt word is optimized according to the evaluation result.

[0172] Therefore, through the evaluation results of the training samples generated by the prompt words, the prompt words can be optimized to make the prompt words more accurate, thereby generating higher quality training samples for the promotion needs corresponding to the prompt words.

[0173] In a possible implementation, the evaluation result includes a professional evaluation result and a promotion effect evaluation result.

[0174] Through multi-faceted evaluation results, we can accurately control the expected effects of product promotion and make targeted adjustments to the prompt words.

[0175] In a possible implementation manner, the determining unit is specifically configured to:

[0176] Based on the attention mechanism model, the professional features and the copywriting style features are fused to obtain fused features;

[0177] The relevant corpus is classified according to the fusion features to determine the multiple style tags.

[0178] This can improve the performance of classification tasks and ensure the accuracy of the copywriting style and product type identification of the relevant corpus.

[0179] In a possible implementation manner, the generating unit is further configured to:

[0180] The evaluation results of the two dimensions of professionalism and promotion effect are analyzed. If the evaluation result of the target dimension does not meet the expected value, the content corresponding to the target dimension in the prompt word is adjusted. The target dimension is any one or both of the two dimensions of professionalism and promotion effect.

[0181] Therefore, by optimizing the prompt words through the evaluation results of the two dimensions, the promotional copy generated according to the training samples can achieve the promotion effect expected by the promoter.

[0182] On the other hand, the present application provides a computer device, characterized in that the computer device includes a processor and a memory:

[0183] The memory is used to store computer programs;

[0184] The processor is used to execute the steps of the method for constructing the promotional copy training corpus according to the computer program.

[0185] On the basis of the above-mentioned embodiments, the embodiments of the present application further provide a computer-readable medium, wherein the computer-readable medium is used to store a computer program, and when the computer program is executed by a computer device, the steps of the method for constructing the above-mentioned promotional copy training corpus are implemented.

[0186] On the basis of the above embodiments, the embodiments of the present application further provide a computer program product including a computer program, which, when executed on a computer device, enables the computer device to execute the steps of the method for constructing a promotional copy training corpus.

[0187] It should be noted that the various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments, and the same or similar parts between the various embodiments can be referred to each other. For the system or device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part description.

[0188] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to the embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for constructing a promotional copy training corpus, characterized in that: The method comprises: Obtain the first training corpus of promotional copy; Extracting professional features and copywriting style features of the first training corpus, wherein the professional features are used to identify the promotion product type of the first training corpus, and the copywriting style features are used to identify the copywriting style of the first training corpus; Classifying the first training corpus according to the professional features and the copywriting style features to determine a plurality of style tags, wherein the style tags are used to indicate a category combination of a promotional product type and a copywriting style; Generate an indicator word according to a target style tag among the multiple style tags, a training corpus corresponding to the target style tag, and professional knowledge corresponding to the target style tag; The indicator word is input into a first generation model to generate a training sample corresponding to the target style label, and the training sample is used to train a promotional copy recognition model, and the promotional copy recognition model is used to recognize the promotional product type and copy style of the input copy to be recognized.

2. The method according to claim 1, characterized in that The expertise includes professional terminology, promotional product information, and compliance content.

3. The method according to claim 1, characterized in that After inputting the indicator word into the first generation model to generate a training sample corresponding to the target style label, the method further includes: Determine reference copy that meets the promotion needs corresponding to the target style tag; Inputting the prompt word into a second generation model to generate text corresponding to the target style label, wherein the scale of the second generation model is smaller than that of the first generation model; Evaluate the document according to the reference document to obtain an evaluation result; The prompt word is optimized according to the evaluation result.

4. The method according to claim 3, characterized in that The evaluation results include professional evaluation results and promotion effect evaluation results.

5. The method according to claim 1, characterized in that: The classifying the first training corpus according to the professional features and the copywriting style features to determine a plurality of style tags includes: Based on the attention mechanism model, the professional features and the copywriting style features are fused to obtain fused features; The relevant corpus is classified according to the fusion features to determine the multiple style tags.

6. The method according to claim 4, characterized in that Optimizing the prompt word according to the evaluation result includes: The evaluation results of the two dimensions of professionalism and promotion effect are analyzed. If the evaluation result of the target dimension does not meet the expected value, the content corresponding to the target dimension in the prompt word is adjusted. The target dimension is any one or both of the two dimensions of professionalism and promotion effect.

7. A device for constructing promotional copy training corpus, characterized in that: The device comprises an acquisition unit, an extraction unit, a determination unit and a generation unit: The acquisition unit is used to acquire a first training corpus of the promotional copy; The extraction unit is used to extract professional features and copywriting style features of the first training corpus, wherein the professional features are used to identify the promotion product type of the first training corpus, and the copywriting style features are used to identify the copywriting style of the first training corpus; The determining unit is used to classify the first training corpus according to the professional characteristics and the copywriting style characteristics, and determine a plurality of style labels, wherein the style labels are used to indicate a category combination of a promotional product type and a copywriting style; The generating unit is used to generate an indicator word according to a target style tag among the multiple style tags, a training corpus corresponding to the target style tag, and professional knowledge corresponding to the target style tag; The generation unit is also used to input the indicator word into the first generation model to generate training corpus corresponding to the target style label. The training sample is used to train the promotional copy recognition model, and the promotional copy recognition model is used to identify the promotion product type and copy style of the input copy to be recognized.

8. A computer device, characterized in that: The computer device comprises a processor and a memory: The memory is used to store computer programs; The processor is configured to execute the method according to any one of claims 1 to 6 according to the computer program.

9. A computer-readable medium, characterized in that The computer-readable medium is used to store a computer program, and when the computer program is executed by a computer device, the method according to any one of claims 1 to 6 is implemented.

10. A computer program product comprising a computer program, characterized in that When the method is executed on a computer device, the computer device is enabled to execute the method according to any one of claims 1 to 6.

Images