Advertisement copy style recognition method and device based on neural network

Abstract

The present invention discloses a method and device for identifying the style of advertisement copy based on a neural network, the method comprising: obtaining a target advertisement copy; inputting the target advertisement copy into a trained style identification network model to determine the copy style corresponding to the target advertisement copy; the style identification network model is obtained by training a training data set including a plurality of training advertisement copies marked as having different copy styles. It can be seen that the present invention can use the trained style identification network model to identify the style of advertisement copy, thereby utilizing the algorithm advantages of the neural network algorithm model to improve the accuracy and efficiency of style identification, thereby providing a data basis for subsequent copy generation.

Description

Technical Field

[0001] The present invention relates to the technical field of natural language processing, and in particular to a method and device for identifying the style of advertising copy based on a neural network. Background Art

[0002] With the development of Internet technology, the forms of online advertisements are becoming more and more complex and diversified. In order to improve the efficiency of generating advertisement content, the existing technology begins to introduce neural network algorithms. However, when generating advertisement content, the existing technology generally manually identifies the promotional style of the advertisement copy, and the manual method has the problems of low efficiency and poor accuracy. It can be seen that the existing technology has defects that need to be solved urgently. Summary of the invention

[0003] The technical problem to be solved by the present invention is to provide a method and device for identifying and determining the style of advertising copy based on a neural network, which can utilize the algorithmic advantages of the neural network algorithm model to improve the accuracy and efficiency of style identification, and thus provide a data basis for subsequent copy generation.

[0004] In order to solve the above technical problems, the first aspect of the present invention discloses a method for identifying the style of advertising copy based on a neural network, the method comprising:

[0005] Get targeted advertising copy;

[0006] The target advertising copy is input into a trained style recognition network model to determine the text style corresponding to the target advertising copy; the style recognition network model is trained by a training data set including a plurality of training advertising copies marked with different text styles.

[0007] As an optional implementation, in the first aspect of the present invention, the style recognition network model includes a word vector conversion layer, a context information extraction layer and a classification layer; the step of inputting the target advertising copy into the trained style recognition network model to determine the copy style corresponding to the target advertising copy includes:

[0008] Inputting the target advertisement text into the word vector conversion layer to obtain the target text word vector;

[0009] Inputting the target copy word vector into the context information extraction layer to obtain a fused copy word vector fused with the context information;

[0010] The fused copywriting word vector is input into the classification layer to obtain the corresponding copywriting style.

[0011] As an optional implementation, in the first aspect of the present invention, the style recognition network model further includes a promotion target marking layer and a position weight calculation layer; before inputting the target copy word vector into the context information extraction layer to obtain a fused copy word vector fused with the context information, the method further includes:

[0012] Inputting the target copy word vector into the promotion target marking layer to mark the promotion target word vectors corresponding to all promotion target texts in the target advertising copy;

[0013] Furthermore, before inputting the fused copywriting word vector into the classification layer to obtain the corresponding copywriting style, the method further includes:

[0014] The fused copy word vector is input into the position weight calculation layer to calculate the vector position weight corresponding to each word vector in the fused copy word vector, and the combination of the products of all the word vectors and the corresponding vector position weights is determined as the fused copy word vector; the vector position weight is inversely proportional to the distance between the word vector and the promotion target word vector.

[0015] As an optional implementation, in the first aspect of the present invention, the style recognition network model further includes an importance weight calculation layer;

[0016] Furthermore, before inputting the fused copywriting word vector into the classification layer to obtain the corresponding copywriting style, the method further includes:

[0017] The fused copy word vector is input into the importance weight calculation layer to calculate the vector importance weight corresponding to each word vector in the fused copy word vector, and the combination of the products of all the word vectors and the corresponding vector importance weights is determined as the fused copy word vector; the vector importance weight includes the word frequency-inverse file frequency of the text corresponding to the word vector, and / or the part-of-speech importance weight of the text corresponding to the word vector.

[0018] As an optional implementation, in the first aspect of the present invention, the step of calculating the importance weight of the part of speech of the text corresponding to the word vector includes:

[0019] Calculate the part-of-speech combination of the text corresponding to the word vector;

[0020] Calculating a frequency parameter of the part-of-speech combination; the frequency parameter is proportional to the number of times the part-of-speech combination appears in the target advertisement copy;

[0021] Calculating an inverse frequency parameter of the part-of-speech combination; the inverse frequency parameter is inversely proportional to the number of advertisements containing the part-of-speech combination in all advertisements;

[0022] The part-of-speech importance weight of the text corresponding to the word vector is determined according to the frequency parameter and the inverse frequency parameter of the part-of-speech combination.

[0023] As an optional implementation, in the first aspect of the present invention, the style recognition network model is trained based on the following steps:

[0024] Determine a plurality of training advertising copies that are labeled as having different copy styles;

[0025] Determine a style recognition training model; the style recognition training model includes the style recognition network model and a training convergence layer; the training convergence layer is used to adjust the model parameters of the style recognition network model during training to minimize the corresponding loss function;

[0026] The plurality of training advertisement texts are input into the style recognition training model for training until convergence, thereby obtaining the trained style recognition network model.

[0027] As an optional embodiment, in the first aspect of the present invention, the context information extraction layer includes at least one of a BiLSTM model, a BiGRU model, a pQRNN model, an ALBERT model, a BERT model, a DistilBERT model, an ELMo model, and a Transformer model.

[0028] As an optional implementation, in the first aspect of the present invention, the style recognition network model further includes a conditional probability calculation layer; the step of inputting the fused copy word vector into the classification layer to obtain the corresponding copy style includes:

[0029] Inputting the fused copy word vector into the conditional probability calculation layer to calculate the conditional probability that each word vector in the fused copy word vector belongs to the promotion feature information;

[0030] Determining the promotion feature information in the target advertisement copy according to the conditional probability that all word vectors in the fused copy word vector belong to the promotion feature information;

[0031] Each word vector in the fused copy word vector is multiplied by the corresponding conditional probability to obtain a fused probability copy word vector, and the fused probability copy word vector is input into the classification layer to obtain the corresponding copy style.

[0032] The second aspect of the present invention discloses a device for identifying the style of advertising copy based on a neural network, the device comprising:

[0033] An acquisition module is used to acquire target advertisement copy;

[0034] The recognition module is used to input the target advertising copy into a trained style recognition network model to determine the copy style corresponding to the target advertising copy; the style recognition network model is trained by a training data set including a plurality of training advertising copies marked with different copy styles.

[0035] As an optional implementation, in the second aspect of the present invention, the style recognition network model includes a word vector conversion layer, a context information extraction layer and a classification layer; the recognition module includes:

[0036] A word vector conversion unit, used for inputting the target advertisement text into the word vector conversion layer to obtain a target text word vector;

[0037] A context fusion unit, used for inputting the target copy word vector into the context information extraction layer to obtain a fused copy word vector fused with the context information;

[0038] The classification unit is used to input the fused copy word vector into the classification layer to obtain the corresponding copy style.

[0039] As an optional implementation, in the second aspect of the present invention, the style recognition network model further includes a promotion target marking layer and a position weight calculation layer; the recognition module further includes:

[0040] a marking unit, configured to input the target copy word vector into the promotion target marking layer before the context fusion unit inputs the target copy word vector into the context information extraction layer to obtain a fused copy word vector fused with the context information, so as to mark the promotion target word vectors corresponding to all promotion target texts in the target advertising copy;

[0041] And, the identification module also includes:

[0042] A position weight calculation unit is used to input the fused copy word vector into the position weight calculation layer before the classification unit inputs the fused copy word vector into the classification layer to obtain the corresponding copy style, so as to calculate the vector position weight corresponding to each word vector in the fused copy word vector, and determine the combination of the products of all the word vectors and the corresponding vector position weights as the fused copy word vector; the vector position weight is inversely proportional to the distance between the word vector and the promotion target word vector.

[0043] As an optional implementation, in the second aspect of the present invention, the style recognition network model further includes an importance weight calculation layer;

[0044] And, the identification module also includes:

[0045] The importance weight calculation unit is used to input the fused copy word vector into the importance weight calculation layer before the classification unit inputs the fused copy word vector into the classification layer to obtain the corresponding copy style, so as to calculate the vector importance weight corresponding to each word vector in the fused copy word vector, and determine the combination of the products of all the word vectors and the corresponding vector importance weights as the fused copy word vector; the vector importance weight includes the word frequency-inverse file frequency of the text corresponding to the word vector, and / or the part-of-speech importance weight of the text corresponding to the word vector.

[0046] As an optional implementation, in the second aspect of the present invention, the device further includes a part-of-speech importance calculation module, which is used to perform the following steps to calculate the part-of-speech importance weight of the text corresponding to the word vector:

[0047] Calculate the part-of-speech combination of the text corresponding to the word vector;

[0048] Calculating a frequency parameter of the part-of-speech combination; the frequency parameter is proportional to the number of times the part-of-speech combination appears in the target advertisement copy;

[0049] Calculating an inverse frequency parameter of the part-of-speech combination; the inverse frequency parameter is inversely proportional to the number of advertisements containing the part-of-speech combination in all advertisements;

[0050] The part-of-speech importance weight of the text corresponding to the word vector is determined according to the frequency parameter and the inverse frequency parameter of the part-of-speech combination.

[0051] As an optional implementation, in the second aspect of the present invention, the device further includes a training module, which is used to perform the following steps to train and obtain the style recognition network model:

[0052] Determine a plurality of training advertising copies that are labeled as having different copy styles;

[0053] Determine a style recognition training model; the style recognition training model includes the style recognition network model and a training convergence layer; the training convergence layer is used to adjust the model parameters of the style recognition network model during training to minimize the corresponding loss function;

[0054] The plurality of training advertisement texts are input into the style recognition training model for training until convergence, thereby obtaining the trained style recognition network model.

[0055] As an optional embodiment, in the second aspect of the present invention, the context information extraction layer includes at least one of a BiLSTM model, a BiGRU model, a pQRNN model, an ALBERT model, a BERT model, a DistilBERT model, an ELMo model, and a Transformer model.

[0056] As an optional implementation, in the second aspect of the present invention, the style recognition network model further includes a conditional probability calculation layer; the classification unit inputs the fused copy word vector into the classification layer to obtain the corresponding copy style, including:

[0057] Inputting the fused copy word vector into the conditional probability calculation layer to calculate the conditional probability that each word vector in the fused copy word vector belongs to the promotion feature information;

[0058] Determining the promotion feature information in the target advertisement copy according to the conditional probability that all word vectors in the fused copy word vector belong to the promotion feature information;

[0059] Each word vector in the fused copy word vector is multiplied by the corresponding conditional probability to obtain a fused probability copy word vector, and the fused probability copy word vector is input into the classification layer to obtain the corresponding copy style.

[0060] The third aspect of the present invention discloses another device for identifying the style of advertisement copy based on a neural network, the device comprising:

[0061] A memory storing executable program code;

[0062] a processor coupled to the memory;

[0063] The processor calls the executable program code stored in the memory to execute part or all of the steps in the method for identifying the style of advertising copy based on a neural network disclosed in the first aspect of an embodiment of the present invention.

[0064] A fourth aspect of an embodiment of the present invention discloses a computer storage medium, which stores computer instructions. When the computer instructions are called, they are used to execute some or all of the steps in the neural network-based advertising copy style recognition method disclosed in the first aspect of the embodiment of the present invention.

[0065] Compared with the prior art, the embodiments of the present invention have the following beneficial effects:

[0066] In an embodiment of the present invention, a target advertisement copy is obtained; the target advertisement copy is input into a trained style recognition network model to determine the copy style corresponding to the target advertisement copy; the style recognition network model is trained by a training data set including a plurality of training advertisement copies marked with different copy styles. It can be seen that the present invention can use the trained style recognition network model to identify the style of the advertisement copy, thereby utilizing the algorithm advantages of the neural network algorithm model to improve the accuracy and efficiency of style recognition, thereby providing a data basis for subsequent copy generation. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0068] Figure 1 It is a flowchart of a method for identifying the style of advertisement copy based on a neural network disclosed in an embodiment of the present invention;

[0069] Figure 2 It is a structural schematic diagram of an advertising copy style recognition device based on a neural network disclosed in an embodiment of the present invention;

[0070] Figure 3 It is a structural schematic diagram of another device for identifying the style of advertisement copy based on a neural network disclosed in an embodiment of the present invention;

[0071] Figure 4 It is a structural schematic diagram of another device for identifying advertising copy style based on a neural network disclosed in an embodiment of the present invention. DETAILED DESCRIPTION

[0072] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings 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.

[0073] The terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish different objects rather than to describe a specific order. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, device, product or end including a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units that are not listed, or may optionally include other steps or units inherent to these processes, methods, products or ends.

[0074] Reference to "embodiments" herein means that a particular feature, structure, or characteristic described in conjunction with the embodiments may be included in at least one embodiment of the present invention. The appearance of the phrase in various places in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment that is mutually exclusive with other embodiments. It is explicitly and implicitly understood by those skilled in the art that the embodiments described herein may be combined with other embodiments.

[0075] The present invention discloses a method and device for identifying the style of advertising copy based on a neural network, which can use a trained style identification network model to identify the style of advertising copy, thereby utilizing the algorithm advantages of the neural network algorithm model to improve the accuracy and efficiency of style identification, thereby providing a data basis for subsequent copy generation. The following is a detailed description.

[0076] Embodiment 1

[0077] See also Figure 1 , Figure 1 1 is a flowchart of a method for identifying the style of advertisement copy based on a neural network disclosed in an embodiment of the present invention. Figure 1 The described method is applied to a document text processing device, which may be a corresponding processing terminal, processing equipment or processing server, and the server may be a local server or a cloud server, which is not limited in the embodiments of the present invention. Figure 1 As shown, the method for identifying the style of advertisement copy based on a neural network may include the following operations:

[0078] 101. Obtain targeted advertising copy.

[0079] Optionally, the target advertising copy may be obtained by crawling the text of a specific network address, by receiving advertising copy input after manual screening, or by obtaining advertising text by means of voice recognition or image recognition, which is not limited by the present invention. Optionally, the target advertising copy may be promotional copy for goods or services of different categories, such as promotional copy used in the introduction page of goods or services or in business activities, or promotional copy used to attract traffic to goods or services, such as article titles or story quotes written by specific media such as WeChat public accounts to attract users, which is not limited by the present invention. Optionally, the language of the target advertising copy is not limited to Chinese, and may also be other grammatically clear languages ​​such as English and French, which is not limited by the present invention.

[0080] 102. Input the target advertising copy into the trained style recognition network model to determine the copy style corresponding to the target advertising copy.

[0081] Specifically, the style recognition network model is trained by a training data set including a plurality of training advertisements labeled with different copywriting styles. Optionally, the style recognition network model may be an existing common classification network model. Optionally, the copywriting style is used to indicate the narrative tone style or wording style used by the advertisement copy in the promotion. For example, the advertisement copy corresponding to the special promotion style may be expressed as: "Limited time sale of famous brand lipsticks, hidden coupons waiting for you to take", while the advertisement copy corresponding to the efficacy expression style may be expressed as "Clinique laser bottle, lightens fine lines, whitens and moisturizes in one bottle".

[0082] It can be seen that the method described in the embodiment of the present invention can use the trained style recognition network model to identify the style of the advertising copy, thereby utilizing the algorithm advantages of the neural network algorithm model to improve the accuracy and efficiency of style recognition, thereby providing a data basis for subsequent copy generation.

[0083] As an optional implementation, the style recognition network model includes a word vector conversion layer, a context information extraction layer and a classification layer. In the above step 102, inputting the target advertising copy into the trained style recognition network model to determine the copy style corresponding to the target advertising copy includes:

[0084] Input the target advertising copy into the word vector conversion layer to obtain the target copy word vector;

[0085] Input the target copy word vector into the context information extraction layer to obtain the fused copy word vector after fusion with the context information;

[0086] Input the fused copywriting word vector into the classification layer to obtain the corresponding copywriting style.

[0087] Optionally, the word vector conversion layer can use a word vector conversion model to convert the target advertising copy. For example, a pre-trained word vector model can be used to convert text to vectors, such as the Tencent word vector model or the Sogou word vector model, or a fixed representation model based on word vectors such as fasttext, word2vec, glove, etc. In a specific solution, the word vector conversion layer uses the 200-dimensional Tencent word vector to initialize the target advertising copy to obtain the target copy word vector.

[0088] Optionally, the context information extraction may be a feature extraction neural network, for example, it may include at least one of the feature extraction neural networks such as BiLSTM model, BiGRU model, pQRNN model, ALBERT model, BERT model, DistilBERT model, ELMo model, Transformer model, etc. Preferably, the context information extraction layer is used to extract the context information of any word vector in the target copy word vector and integrate it into the word vector to obtain a fused copy word vector.

[0089] Optionally, the classification layer includes a fully connected layer and a softmax layer, which can map the input fused copy word vector to obtain a probability distribution for different copy styles, that is, output the probability that the target advertising copy belongs to different copy style labels, and select the style label with the maximum probability as the final result.

[0090] It can be seen that implementing this optional implementation method can utilize the word vector conversion layer, context information extraction layer and classification layer to classify the style of the target advertising copy, thereby improving the accuracy and efficiency of style recognition, and further providing a data basis for subsequent copy generation.

[0091] As an optional implementation, the style recognition network model also includes a promotion target tag layer. Accordingly, in the above steps, before inputting the target copy word vector into the context information extraction layer to obtain the fused copy word vector fused with the context information, the method also includes:

[0092] The target copy word vector is input into the promotion target labeling layer to label the promotion target word vector corresponding to all promotion target texts in the target advertising copy.

[0093] Optionally, the promotion target labeling layer labels the promotion target word vector in a manner that generates an indicator vector for all word vectors in the target copy word vector, wherein only the indicator vector of the promotion target word vector is a preset or random value, and the indicator vectors of other word vectors are 0 or other null values, and the indicator vector is concatenated with the word vector to form a new labeled target copy word vector. Optionally, the promotion target is used to characterize the promotion object in the target advertising copy, such as the goods, services or platforms that the target advertising copy intends to promote.

[0094] It can be seen that implementing this optional implementation method can input the target copy word vector into the promotion target marking layer to mark the promotion target word vector corresponding to all promotion target texts in the target advertising copy, thereby accurately locating the promotion target in the target advertising copy.

[0095] As an optional implementation, the style recognition network model also includes a position weight calculation layer. Accordingly, in the above steps, before inputting the fused copy word vector into the classification layer to obtain the corresponding copy style, the method also includes:

[0096] The fused copy word vector is input into the position weight calculation layer to calculate the vector position weight corresponding to each word vector in the fused copy word vector, and the combination of the product of all word vectors and the corresponding vector position weight is determined as the fused copy word vector. Among them, the vector position weight is inversely proportional to the distance between the word vector and the promotion target word vector.

[0097] Optionally, the vector position weight can be calculated according to a piecewise function. Optionally, the piecewise function can be used to limit the vector position weight corresponding to the word vector to a preset fixed value when the word text corresponding to the word vector is in the promotion target text corresponding to the promotion target word vector. Optionally, the fixed value can be 1. When the word text corresponding to the word vector is not in the promotion target text corresponding to the promotion target word vector, the vector position weight corresponding to the word vector is inversely proportional to the distance between the word text and the promotion target text, such as an exponential inverse relationship.

[0098] It can be seen that implementing this optional implementation method can input the fused copy word vector into the position weight calculation layer to calculate the vector position weight corresponding to each word vector in the fused copy word vector, so as to determine the degree of influence of word vectors in different positions on the promotion target, so that when subsequently identifying the style of the advertising copy, the degree of influence of the word vector on the promotion target can be fully considered, so as to fully establish the relationship between the promotion target and the style.

[0099] As an optional implementation, the style recognition network model further includes an importance weight calculation layer. Accordingly, in the above steps, before inputting the fused copy word vector into the classification layer to obtain the corresponding copy style, the method further includes:

[0100] The fused copy word vector is input into the importance weight calculation layer to calculate the vector importance weight corresponding to each word vector in the fused copy word vector, and the combination of the product of all word vectors and the corresponding vector importance weight is determined as the fused copy word vector. Among them, the vector importance weight includes the term frequency-inverse document frequency of the text corresponding to the word vector, and / or the part-of-speech importance weight of the text corresponding to the word vector. Among them, the term frequency-inverse document frequency is the TF-IDF (TermFrequency-Inverse Document Frequency) parameter.

[0101] Optionally, the vector importance weight may be the product of the word frequency-inverse document frequency of the text corresponding to the word vector and the part-of-speech importance weight. Optionally, the steps for calculating the part-of-speech importance weight of the text corresponding to the word vector include:

[0102] Calculate the part-of-speech combination of the text corresponding to the word vector;

[0103] Calculate the frequency parameters of part-of-speech combinations;

[0104] Calculate the inverse frequency parameters of part-of-speech combinations;

[0105] According to the frequency parameter and inverse frequency parameter of the part-of-speech combination, the part-of-speech importance weight of the text corresponding to the word vector is determined.

[0106] Optionally, the frequency parameter is proportional to the number of times the part-of-speech combination appears in the target advertising copy, and the inverse frequency parameter is inversely proportional to the number of copies in which the part-of-speech combination exists in all advertising copies. Preferably, the frequency parameter can be the ratio of the number of times the part-of-speech combination appears in the target advertising copy to the number of all part-of-speech combinations in the target advertising copy. Preferably, the inverse frequency parameter is the ratio of the total number of all advertising copies to the number of copies in which the part-of-speech combination exists. Among them, all advertising copies include the target advertising copy, and all advertising copies can be all advertising copies in historical training or in the current database. Optionally, the part-of-speech importance weight of the text corresponding to the word vector can be determined as the product of the frequency parameter of the part-of-speech combination and the inverse frequency parameter.

[0107] It can be seen that implementing this optional implementation method can input the fused copy word vector into the importance weight calculation layer to calculate the vector importance weight corresponding to each word vector in the fused copy word vector, so that the importance of different word vectors can be determined based on the frequency information of the word vector and the frequency information of the part-of-speech combination, so that when subsequently identifying the style of the advertising copy, the influence of the importance of the word vector on the style can be fully considered to improve the accuracy of style recognition.

[0108] As an optional implementation, the style recognition network model may include at least two of a promotion target marking layer, a position weight calculation layer, and an importance weight calculation layer. Correspondingly, the method may also include a combination of at least two of the above three optional implementations, and the order of the steps is not limited.

[0109] As an optional implementation, the style recognition network model is trained based on the following steps:

[0110] Determine a plurality of training advertising copies that are labeled as having different copy styles;

[0111] Determine the style recognition training model;

[0112] Multiple training advertisements are input into the style recognition training model for training until convergence, thereby obtaining a trained style recognition network model.

[0113] Among them, the style recognition training model includes a style recognition network model and a training convergence layer. Specifically, the training convergence layer is used to adjust the model parameters of the style recognition network model during training to minimize the corresponding loss function. Optionally, the network layer in the style recognition network model can be the network layer content described in any of the above multiple implementations. Optionally, the loss function can be a minimized negative log-likelihood function, which is used to calculate the distance between the style classification result of the training advertising copy by the classification layer and the annotated style, and the training convergence layer is used to continuously adjust the model parameters of different network layers of the style recognition network model so that the final output style is as close to the annotated style as possible.

[0114] It can be seen that implementing this optional implementation method can input multiple training advertising texts into the style recognition training model for training until convergence, thereby obtaining a trained style recognition network model, thereby being able to train a style recognition network model with a higher recognition rate. The style recognition network model can be used in the subsequent process to improve the accuracy and efficiency of style recognition, thereby providing a data basis for subsequent text generation.

[0115] As an optional implementation, the style recognition network model further includes a conditional probability calculation layer. Accordingly, in the above steps, the fused text word vector is input into the classification layer to obtain the corresponding text style, including:

[0116] Input the fused copy word vector into the conditional probability calculation layer to calculate the conditional probability that each word vector in the fused copy word vector belongs to the promotion feature information;

[0117] Determine the promotion feature information in the target advertising copy based on the conditional probability that all word vectors in the fused copy word vector belong to the promotion feature information;

[0118] Multiply each word vector in the fused copy word vector by the corresponding conditional probability to obtain the fused probability copy word vector, and input the fused probability copy word vector into the classification layer to obtain the corresponding copy style.

[0119] Optionally, the promotion feature information may include at least one of the following: promotion object category information, promotion object ingredient information, promotion object efficacy information, promotion object brand information, and promotional marketing term information. For example, "mobile phone screen film" belongs to promotion object category information, while "high performance", "foldable design", "anti-fingerprint" or "anti-scratch" all belong to promotion object efficacy information, and "large coupons" and "limited time flash sales" belong to promotional marketing term information.

[0120] Optionally, the text segments corresponding to all word vectors whose conditional probability of belonging to promotion feature information in the fused copy word vector is higher than a preset probability threshold and / or ranked in the top preset number of places from high to low can be determined as the promotion feature information in the target advertising copy.

[0121] The purpose of the above improvement is that different advertising copywriting styles contain different promotional feature information. For example, the special promotion style will pay more attention to marketing term information, while the efficacy expression style will focus on efficacy ingredient information. Different promotional goals, such as advertising copywriting for different product categories, also contain different promotional feature information. For example, some virtual service-type advertising copywriting, such as mobile game recharge platforms, will pay more attention to marketing term information. The copywriting of some medical supplies will pay more attention to ingredient information and efficacy information regardless of the copywriting style. Therefore, by introducing the conditional probability calculation layer, on the one hand, the promotional feature information in the target advertising copy can be identified, and on the other hand, the corresponding copywriting style can be identified by using the word vector after the fusion of the promotion target information and the conditional probability, so as to establish the connection between the copywriting style and the promotional feature information, as well as the connection between the promotion target and the promotional feature information, so as to train a style recognition network model that can simultaneously identify the promotion target, copywriting style and promotional feature information.

[0122] As an optional implementation, the conditional probability calculation layer includes multiple conditional probability calculation models, and in the above steps, the fused copy word vector is input into the conditional probability calculation layer to calculate the conditional probability that each word vector in the fused copy word vector belongs to the promotion feature information, including:

[0123] Input the fused copy word vectors into each conditional probability calculation model in parallel to obtain the conditional probability that each word vector in the fused copy word vector calculated by each conditional probability calculation model belongs to the promotion feature information;

[0124] For any word vector in the fused copy word vector, the conditional probability that the word vector belongs to the promotion feature information is calculated based on all conditional probability calculation models to determine the conditional probability that the word vector belongs to the promotion feature information.

[0125] Optionally, the conditional probability that the word vector belongs to the generalized feature information may be determined as the conditional probability that the word vector belongs to the generalized feature information, which is calculated by all conditional probability calculation models or an added value.

[0126] The purpose of the above improvement is to calculate the conditional probability that the word vector belongs to the promotion feature information in parallel through multiple conditional probability calculation models, so that multiple conditional probability calculation models can pay attention to different promotion feature information in the fused copy word vector, and combine the conditional probabilities obtained by multiple conditional probability calculation models to accurately determine the promotion feature information.

[0127] As an optional implementation, in the above step, multiplying each word vector in the fused copy word vector by the corresponding conditional probability to obtain the fused probability copy word vector includes:

[0128] Multiply each word vector in the fused copy word vector by the conditional probability calculated by each conditional probability calculation model to obtain multiple sub-fused probability copy word vectors;

[0129] Multiple sub-fusion probability copy word vectors are concatenated to obtain a fusion probability copy word vector.

[0130] Preferably, the conditional probability calculation model is a CRF (Conditional Random Fields) model.

[0131] The purpose of the above improvement is to obtain a fused probabilistic copywriting word vector by splicing multiple sub-fused probabilistic copywriting word vectors. When subsequently performing style prediction based on the fused probabilistic copywriting word vector, the promotion feature information can be considered during style classification, so that the final style classification result can be associated with the promotion feature information.

[0132] Accordingly, in this optional implementation, the style recognition network model is trained using a plurality of training advertisement texts that are annotated with text style and promotion feature information. Optionally, the style recognition network model is trained based on the following steps:

[0133] Determine a plurality of training advertisement copies that are marked with copywriting styles and promotion feature information; wherein the promotion feature information is related to the copywriting style;

[0134] Determine a feature information recognition training model; wherein the feature information recognition training model includes a style recognition network model and a training convergence layer; the training convergence layer is used to adjust the model parameters of the conditional probability calculation layer and the classification layer in the style recognition network model during training to minimize the loss function corresponding to the conditional probability calculation layer and the classification layer;

[0135] A plurality of training advertisement copies are input into the feature information recognition training model for training until convergence, thereby obtaining a trained style recognition network model.

[0136] Optionally, the copywriting style and promotion feature information of the training advertisement copy can be annotated with a copywriting style label for each training copy, and a promotion feature information label can be created for each text in the training advertisement copy to mark whether each context word in the training copy belongs to the promotion feature information of the advertisement copy, such as the category, brand, efficacy, ingredient, marketing term fragment, etc. The purpose of this step is to mark the promotion feature information of the advertisement copy. At the same time, the promotion feature information is related to the copywriting style, so when annotating, it is necessary to follow the emphasis of different copywriting styles. For example, for the training advertisement copy of the copywriting style of efficacy description, all the efficacy words in it need to be annotated as promotion feature information, and other words need to be annotated as irrelevant information. If the training advertisement copy belongs to the copywriting style of promotion-oriented copywriting, all the marketing terms in it need to be annotated as promotion feature information, and other words need to be annotated as irrelevant information. The purpose of this is to manually define the promotion feature information corresponding to different copywriting styles, so that the model can learn it later.

[0137] Optionally, when the feature information identification training model includes a conditional probability calculation layer composed of multiple CRF models, the training convergence layer is also used to minimize the loss function corresponding to each CRF model so that the feature information identification training model converges. Optionally, the conditional probability calculation layer also includes an encoding layer, which is used to map the input fused copy word vector to obtain a first probability that each word vector in the fused copy word vector belongs to the promotion feature information and a second probability that it does not belong to the promotion feature information, and input the fused copy word vector including the first probability and the second probability into multiple CRF models. Optionally, each CRF model will calculate the emission score corresponding to each word vector in the fused copy word vector based on the first probability and the second probability, and calculate the transfer score of any two adjacent word vectors in the fused copy word vector based on the trained transfer matrix parameters, and calculate the conditional probability that each word vector in the fused copy word vector in the correct path belongs to the promotion feature information based on all the emission scores and the transfer scores.

[0138] Optionally, when training the feature information recognition training model, each CRF model will randomly generate transfer matrix parameters and continuously optimize them during training. The loss function corresponding to the CRF model is the maximum likelihood loss function, which is used to calculate the distance between the sum of the correct path scores calculated by each CRF model and the sum of all path scores, where the sum of the correct path scores is the sum of all emission scores and transfer scores corresponding to the fused copy word vector when the CRF model calculates the correct path. The correct path is a path representation that conforms to the promotion feature information annotation label in the training copy.

[0139] Optionally, each CRF model can calculate the marginal distribution of the promotion feature information label, and multiply the marginal distribution of the promotion feature information label by the fused copy word vector to obtain the above-mentioned sub-fused probability copy word vector, and then concatenate the sub-fused probability copy word vectors output by all CRF models to obtain the fused probability copy word vector.

[0140] As can be seen from the above, the above-mentioned style recognition network model of the present invention uses multiple CRF models in parallel, which can achieve different methods of obtaining feature subsets, respectively modeling the corresponding feature subsets, forming multiple CRF models for feature subsets, and then combining the results of all CRF models, which is equivalent to feature selection, to capture the important promotion feature information points corresponding to the advertising focus target. The purpose of this setting is to use a method similar to the attention mechanism to measure the importance of different promotion feature information to the style of the promotional copy. It is proved in the experiment that different CRF models pay attention to different promotion feature information, so that the output of multiple CRF models can be combined to derive the different importance of the important promotion feature information points of the copy to the advertising copy style under the promotion target. This part also has the function of extracting promotion feature information. After the model is trained, the important promotion feature information of each advertising text is also output through the CRF model.

[0141] It can be seen that the advertising copy style recognition method described in the present invention can realize the extraction of promotion feature information points in advertising copy with different copy styles under various promotion goals, and identify the copy style of the copy, thereby establishing the relationship between the two dimensions of promotion feature information point-advertising copy style and promotion feature information point-advertising promotion object. At the same time, the advertising copy feature information recognition of this method not only focuses on the recommended object of the advertising text, but also classifies the writing method and structure of the advertising copy. It can be used for deep-level advertising writing style mining, and can take into account the order relationship between words and the combination and arrangement of different words to achieve better mining effects. Optionally, this method can also be used to automatically output a list of promotion feature information after inputting the user's promotion goals and copy style. This can be achieved by data statistics, and is conducive to providing a data basis for the subsequent automatic generation of copy.

[0142] Embodiment 2

[0143] See also Figure 2 , Figure 2 : is a schematic diagram of a structure of a device for identifying the style of advertisement copy based on a neural network disclosed in an embodiment of the present invention. Figure 2 The described device can be applied to a corresponding document text processing device, which can be a corresponding processing terminal, processing equipment or processing server, and the server can be a local server or a cloud server, which is not limited in the embodiment of the present invention. Figure 2 As shown, the device may include:

[0144] An acquisition module 201 is used to acquire a target advertisement copy;

[0145] The identification module 202 is used to input the target advertising copy into the trained style identification network model to determine the copy style corresponding to the target advertising copy.

[0146] Specifically, the style recognition network model is trained by using a training data set including a plurality of training advertisement texts labeled with different text styles. Optionally, the style recognition network model may be an existing common classification network model.

[0147] It can be seen that the device described in the embodiment of the present invention can use the trained style recognition network model to identify the style of the advertising copy, thereby utilizing the algorithm advantages of the neural network algorithm model to improve the accuracy and efficiency of style recognition, thereby providing a data basis for subsequent copy generation.

[0148] As an optional implementation, the style recognition network model includes a word vector conversion layer, a context information extraction layer and a classification layer, such as Figure 3 As shown, the identification module 202 includes:

[0149] The word vector conversion unit 2021 is used to input the target advertisement text into the word vector conversion layer to obtain the target text word vector;

[0150] The context fusion unit 2022 is used to input the target copy word vector into the context information extraction layer to obtain a fused copy word vector fused with the context information;

[0151] The classification unit 2023 is used to input the fused copy word vector into the classification layer to obtain the corresponding copy style.

[0152] As an optional implementation, the style recognition network model also includes a promotion target marking layer and a position weight calculation layer, such as Figure 3 As shown, the identification module 202 also includes:

[0153] The marking unit 2024 is used to input the target copy word vector into the promotion target marking layer before the context fusion unit 2022 inputs the target copy word vector into the context information extraction layer to obtain the fused copy word vector fused with the context information, so as to mark the promotion target word vector corresponding to all promotion target texts in the target advertising copy;

[0154] And, if Figure 3 As shown, the identification module 202 also includes:

[0155] The position weight calculation unit 2025 is used to input the fused copy word vector into the position weight calculation layer before the classification unit 2023 inputs the fused copy word vector into the classification layer to obtain the corresponding copy style, so as to calculate the vector position weight corresponding to each word vector in the fused copy word vector, and determine the combination of the products of all word vectors and the corresponding vector position weights as the fused copy word vector; the vector position weight is inversely proportional to the distance between the word vector and the promotion target word vector.

[0156] As an optional implementation, the style recognition network model further includes an importance weight calculation layer;

[0157] And, if Figure 4 As shown, the identification module 202 also includes:

[0158] The importance weight calculation unit 2026 is used to input the fused copy word vector into the importance weight calculation layer before the classification unit 2023 inputs the fused copy word vector into the classification layer to obtain the corresponding copy style, so as to calculate the vector importance weight corresponding to each word vector in the fused copy word vector, and determine the combination of the products of all word vectors and the corresponding vector importance weights as the fused copy word vector; the vector importance weight includes the word frequency-inverse file frequency of the text corresponding to the word vector, and / or the part-of-speech importance weight of the text corresponding to the word vector.

[0159] As an optional implementation, Figure 3 As shown, the device also includes a part-of-speech importance calculation module 203, which is used to perform the following steps to calculate the part-of-speech importance weight of the text corresponding to the word vector:

[0160] Calculate the part-of-speech combination of the text corresponding to the word vector;

[0161] Calculate the frequency parameter of the part-of-speech combination; the frequency parameter is proportional to the number of times the part-of-speech combination appears in the target advertising copy;

[0162] Calculate the inverse frequency parameter of the part-of-speech combination; the inverse frequency parameter is inversely proportional to the number of advertisements containing the part-of-speech combination in all advertisements;

[0163] According to the frequency parameter and inverse frequency parameter of the part-of-speech combination, the part-of-speech importance weight of the text corresponding to the word vector is determined.

[0164] As an optional implementation, Figure 3 As shown, the device further includes a training module 204, which is used to perform the following steps to train and obtain a style recognition network model:

[0165] Determine a plurality of training advertising copies that are labeled as having different copy styles;

[0166] Determine a style recognition training model; the style recognition training model includes a style recognition network model and a training convergence layer; the training convergence layer is used to adjust the model parameters of the style recognition network model during training to minimize the corresponding loss function;

[0167] Multiple training advertisements are input into the style recognition training model for training until convergence, thereby obtaining a trained style recognition network model.

[0168] As an optional implementation, the style recognition network model further includes a conditional probability calculation layer; the classification unit 2023 inputs the fused copy word vector into the classification layer to obtain the corresponding copy style, including:

[0169] Input the fused copy word vector into the conditional probability calculation layer to calculate the conditional probability that each word vector in the fused copy word vector belongs to the promotion feature information;

[0170] Determine the promotion feature information in the target advertising copy based on the conditional probability that all word vectors in the fused copy word vector belong to the promotion feature information;

[0171] Multiply each word vector in the fused copy word vector by the corresponding conditional probability to obtain the fused probability copy word vector, and input the fused probability copy word vector into the classification layer to obtain the corresponding copy style.

[0172] The text style recognition device described in this embodiment is a device for executing part or all of the steps of the text style recognition method described in Example 1. Its specific technical details and optional implementation methods can refer to the description of Example 1, and will not be repeated in this embodiment.

[0173] Embodiment 3

[0174] See also Figure 4 , Figure 4 FIG. 1 is a schematic diagram of the structure of another device for identifying the style of advertisement text based on a neural network disclosed in an embodiment of the present invention. Figure 4 As shown, the device may include:

[0175] A memory 301 storing executable program codes;

[0176] a processor 302 coupled to the memory 301;

[0177] The processor 302 calls the executable program code stored in the memory 301 to execute part or all of the steps in the method for identifying the style of advertisement copy based on a neural network disclosed in the first embodiment of the present invention.

[0178] Embodiment 4

[0179] An embodiment of the present invention discloses a computer storage medium, which stores computer instructions. When the computer instructions are called, they are used to execute some or all of the steps in the method for identifying the style of advertising copy based on a neural network disclosed in the first embodiment of the present invention.

[0180] The device embodiments described above are only illustrative, wherein the modules described as separate components may or may not be physically separated, and the components displayed as modules may or may not be physical modules, i.e., they may be located in one place, or they may be distributed on multiple network modules. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Those of ordinary skill in the art may understand and implement it without creative work.

[0181] Through the specific description of the above embodiments, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on such an understanding, the above technical solution can be essentially or partly contributed to the prior art in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, and the storage medium includes a read-only memory (ROM), a random access memory (RAM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), a one-time programmable read-only memory (OTPROM), an electronically erasable rewritable read-only memory (EEPROM), a compact disc (CD-ROM) or other optical disc storage, a magnetic disk storage, a magnetic tape storage, or any other computer-readable medium that can be used to carry or store data.

[0182] Finally, it should be noted that the method and device for advertising copy style recognition based on neural network disclosed in the embodiment of the present invention only disclose the preferred embodiments of the present invention, which are only used to illustrate the technical solution of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, it should be understood by those skilled in the art that the technical solutions described in the aforementioned embodiments can still be modified, or some of the technical features therein can be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A method for identifying advertising copy style based on a neural network, characterized in that: The method comprises: Get targeted advertising copy; Inputting the target advertisement copy into a trained style recognition network model to determine the copy style corresponding to the target advertisement copy; the style recognition network model is trained by a training data set including a plurality of training advertisement copies marked as having different copy styles; The style recognition network model includes a word vector conversion layer, a context information extraction layer and a classification layer; the step of inputting the target advertising copy into the trained style recognition network model to determine the copy style corresponding to the target advertising copy includes: Inputting the target advertisement text into the word vector conversion layer to obtain the target text word vector; Inputting the target copy word vector into the context information extraction layer to obtain a fused copy word vector fused with the context information; Inputting the fused copywriting word vector into the classification layer to obtain the corresponding copywriting style; Furthermore, the style recognition network model further includes a promotion target marking layer and a position weight calculation layer; before inputting the target copy word vector into the context information extraction layer to obtain a fused copy word vector fused with the context information, the method further includes: Inputting the target copy word vector into the promotion target marking layer to mark the promotion target word vectors corresponding to all promotion target texts in the target advertising copy; Furthermore, before inputting the fused copywriting word vector into the classification layer to obtain the corresponding copywriting style, the method further includes: The fused copy word vector is input into the position weight calculation layer to calculate the vector position weight corresponding to each word vector in the fused copy word vector, and the combination of the products of all the word vectors and the corresponding vector position weights is determined as the fused copy word vector; the vector position weight is inversely proportional to the distance between the word vector and the promotion target word vector.

2. The method for identifying advertising copy style based on a neural network according to claim 1, characterized in that: The style recognition network model also includes an importance weight calculation layer; Furthermore, before inputting the fused copywriting word vector into the classification layer to obtain the corresponding copywriting style, the method further includes: The fused copy word vector is input into the importance weight calculation layer to calculate the vector importance weight corresponding to each word vector in the fused copy word vector, and the combination of the products of all the word vectors and the corresponding vector importance weights is determined as the fused copy word vector; the vector importance weight includes the word frequency-inverse file frequency of the text corresponding to the word vector, and / or the part-of-speech importance weight of the text corresponding to the word vector.

3. The method for identifying advertising copy style based on a neural network according to claim 2, characterized in that: The steps of calculating the importance weight of the part of speech of the text corresponding to the word vector include: Calculate the part-of-speech combination of the text corresponding to the word vector; Calculating a frequency parameter of the part-of-speech combination; the frequency parameter is proportional to the number of times the part-of-speech combination appears in the target advertisement copy; Calculating an inverse frequency parameter of the part-of-speech combination; the inverse frequency parameter is inversely proportional to the number of advertisements containing the part-of-speech combination in all advertisements; The part-of-speech importance weight of the text corresponding to the word vector is determined according to the frequency parameter and the inverse frequency parameter of the part-of-speech combination.

4. The method for identifying advertising copy style based on a neural network according to any one of claims 1 to 3, characterized in that: The style recognition network model is trained based on the following steps: Determine a plurality of training advertising copies that are labeled as having different copy styles; Determine a style recognition training model; the style recognition training model includes the style recognition network model and a training convergence layer; the training convergence layer is used to adjust the model parameters of the style recognition network model during training to minimize the corresponding loss function; The plurality of training advertisement texts are input into the style recognition training model for training until convergence, thereby obtaining the trained style recognition network model.

5. The method for identifying advertising copy style based on a neural network according to claim 1, characterized in that: The context information extraction layer includes at least one of a BiLSTM model, a BiGRU model, a pQRNN model, an ALBERT model, a BERT model, a DistilBERT model, an ELMo model, and a Transformer model.

6. The method for identifying advertising copy style based on a neural network according to claim 1, characterized in that: The style recognition network model further includes a conditional probability calculation layer; the step of inputting the fused copy word vector into the classification layer to obtain the corresponding copy style includes: Inputting the fused copy word vector into the conditional probability calculation layer to calculate the conditional probability that each word vector in the fused copy word vector belongs to the promotion feature information; Determining the promotion feature information in the target advertisement copy according to the conditional probability that all word vectors in the fused copy word vector belong to the promotion feature information; Each word vector in the fused copy word vector is multiplied by the corresponding conditional probability to obtain a fused probability copy word vector, and the fused probability copy word vector is input into the classification layer to obtain the corresponding copy style.

7. A device for identifying the style of advertising copy based on a neural network, characterized in that: The device comprises: An acquisition module is used to acquire target advertisement copy; an identification module, configured to input the target advertisement copy into a trained style identification network model to determine the copy style corresponding to the target advertisement copy; the style identification network model is trained by a training data set including a plurality of training advertisement copies marked as having different copy styles; The style recognition network model includes a word vector conversion layer, a context information extraction layer and a classification layer; the recognition module includes: A word vector conversion unit, used for inputting the target advertisement text into the word vector conversion layer to obtain a target text word vector; A context fusion unit, used for inputting the target copy word vector into the context information extraction layer to obtain a fused copy word vector fused with the context information; A classification unit, used for inputting the fused copywriting word vector into the classification layer to obtain a corresponding copywriting style; Furthermore, the style recognition network model further includes a promotion target marking layer and a position weight calculation layer; the recognition module further includes: a marking unit, configured to input the target copy word vector into the promotion target marking layer before the context fusion unit inputs the target copy word vector into the context information extraction layer to obtain a fused copy word vector fused with the context information, so as to mark the promotion target word vectors corresponding to all promotion target texts in the target advertising copy; And, the identification module also includes: A position weight calculation unit is used to input the fused copy word vector into the position weight calculation layer before the classification unit inputs the fused copy word vector into the classification layer to obtain the corresponding copy style, so as to calculate the vector position weight corresponding to each word vector in the fused copy word vector, and determine the combination of the products of all the word vectors and the corresponding vector position weights as the fused copy word vector; the vector position weight is inversely proportional to the distance between the word vector and the promotion target word vector.

8. A device for identifying the style of advertising copy based on a neural network, characterized in that: The device comprises: A memory storing executable program code; a processor coupled to the memory; The processor calls the executable program code stored in the memory to execute the advertising copy style recognition method based on neural network as described in any one of claims 1-6.

Images