Advertisement content generation method and system based on natural language processing

Abstract

The application belongs to the technical field of natural language processing and computing advertising, and discloses an advertisement content generation method and system based on natural language processing. By constructing a creative semantic orthogonal decoupling space and generating an advertisement script, audience cognitive response data is collected to establish a coupling relationship model between creativity and audience cognition, resonance source semantic attribute components that produce a coupling resonance effect with the audience cognition dimension are identified based on a system stability criterion value, and after directional inhibition is applied to the corresponding generation sub-path in the creativity strategy network, the advertisement script is iteratively generated. The application reconstructs the advertisement creativity optimization as a stability control problem of a creativity-audience cognition coupling system, eliminates the resonance instability state of the creativity-audience cognition coupling system, and improves the long-term output stability of the script generated by the creativity strategy network.

Description

Technical Field

[0001] This invention relates to the fields of natural language processing and computational advertising technology, and in particular to a method and system for generating advertising content based on natural language processing. Background Technology

[0002] With the development of natural language processing technology, automatic generation of advertising copy based on large language models has become a core method in the field of advertising technology. Existing technologies construct semantic spaces, deconstruct advertising creatives in multiple dimensions, map the contextual signals of the placement scenario into creative feature vectors, and use audience behavior feedback data to establish a correlation model between creative features and conversion effects, forming an iterative closed loop of "generation-placement-feedback-optimization" to improve the click-through rate and conversion rate of advertising copy.

[0003] However, existing technologies model the relationship between creativity and the audience as a one-way decay process, triggering adjustments by monitoring whether the audience's cognitive response data exceeds a preset threshold. This approach fails to recognize the coupling resonance effect that occurs between the specific semantic attributes of the creative content generated by the natural language processing model and the audience's deep psychological schema—specific copywriting features and specific audience psychological traits form a mutually reinforcing cycle, leading to a continuous decline in the stability of the creative-audience interaction system. This causes divergent oscillations between the output generated by the creative strategy network and the audience's cognitive state, preventing the system from converging to a stable matching state. Existing threshold-triggered circuit breakers or overall copy replacement methods cannot identify the root cause of the coupling resonance effect, nor can they maintain the other semantic dimensions of the copy generated by the creative strategy network while selectively suppressing the specific creative semantic dimensions that cause system instability, resulting in an inability to restore system stability.

[0004] Therefore, this invention proposes a method and system for generating advertising content based on natural language processing. Summary of the Invention

[0005] This invention provides an advertising content generation method and system based on natural language processing. It reconstructs advertising creative optimization into a stability control problem of the creative-audience cognition coupling system. By identifying resonance sources through system stability criteria and suppressing the generated sub-paths, the coupling resonance effect is eliminated from the root, so that the generated output of the creative strategy network and the audience's cognitive state can be restored to stable convergence. This improves the long-term output stability of the copy while keeping other semantic dimensions unchanged.

[0006] This invention provides a method for generating advertising content based on natural language processing, comprising: Construct a creative semantic orthogonal decoupling space, which contains multiple orthogonal dimensions, each corresponding to an independent creative semantic attribute; deploy a creative strategy network to map the context signals of the advertising scenario to creative semantic attribute components in the creative semantic orthogonal decoupling space, and generate the first ad copy for delivery; After the first ad copy is delivered, the audience's response data in multiple cognitive dimensions is obtained. Based on the relationship between the creative semantic attribute components and the audience's response data in multiple cognitive dimensions, a coupling relationship model between the creative semantic attribute components and the audience's response data in multiple cognitive dimensions is established. In the coupling relationship model, based on the system stability criterion, creative semantic attribute components that have a coupling resonance effect with the audience's response data in multiple cognitive dimensions are identified as the resonance source semantic attribute components. In response to a new ad delivery request, the creative semantic attribute generation sub-path corresponding to the semantic attribute component of the resonance source is suppressed, and the creative strategy network generates a second ad copy for delivery.

[0007] Furthermore, a creative semantic orthogonal decoupling space is constructed, specifically including: Multiple orthogonal dimensions are defined, each corresponding to an independent creative semantic attribute. Creative semantic attributes include appeal type, emotional tone, language style, rhetorical devices, and narrative structure. Construct an orthogonal basis vector set, in which any two basis vectors are orthogonal to each other, and each basis vector corresponds to an orthogonal dimension; The creative semantic attribute components are generated by the creative policy network in the creative semantic orthogonal decoupled space spanned by the orthogonal basis vectors. The projection value of the creative semantic attribute component on each orthogonal dimension represents the activation intensity of the corresponding creative semantic attribute.

[0008] Furthermore, after the initial ad copy was placed, the audience's response data across multiple cognitive dimensions was obtained, specifically including: The audience cognitive state space is constructed, which includes four cognitive dimensions: cognitive novelty seeking degree, cognitive fatigue accumulation degree, social identity sensitivity degree, and anchoring effect deviation degree. These four cognitive dimensions together constitute a four-dimensional representation of the audience's psychological schema characteristics. After the first ad copy is placed, collect user behavior data generated by the audience within a preset time window. User behavior data includes page dwell time, scroll depth, hotspot distribution of clicks, and number of times the ad was viewed. Feature extraction is performed on user behavior data to obtain user behavior feature vectors; the user behavior feature vectors are input into a preset behavior-cognition mapping model to calculate the response value corresponding to each cognitive dimension; the response values ​​corresponding to the four cognitive dimensions are combined to form the audience's response data on multiple cognitive dimensions.

[0009] Furthermore, based on the relationship between creative semantic attribute components and audience response data across multiple cognitive dimensions, a coupling relationship model between creative semantic attribute components and audience response data across multiple cognitive dimensions is established, specifically including: Obtain market feedback sequence data generated after the historical ad copy was launched. The historical ad copy refers to the ad copy that was launched before the first ad copy was launched. Combine the creative semantic attribute components of the historical ad copy with the audience's response data in four cognitive dimensions during the corresponding campaign period to form input-output sequence pairs. The input-output sequence pairs are systematically identified, and the gain coefficient, natural frequency, and damping ratio of the transfer function between each pair of creative semantic attribute components and the audience's response data in one of the four cognitive dimensions are estimated. The transfer function adopts the form of a second-order oscillatory element. The transfer function matrix is ​​composed of transfer functions. The element in the i-th row and j-th column of the transfer function matrix is ​​the transfer function between the i-th creative semantic attribute component and the j-th cognitive dimension. The transfer function matrix serves as a coupling relationship model between the creative semantic attribute component and the audience's response data in the four cognitive dimensions.

[0010] Furthermore, in the coupling relationship model, based on the system stability criterion, creative semantic attribute components that exhibit coupling resonance effects with the audience's response data across multiple cognitive dimensions are identified. These components specifically include: Using the creative strategy network as the forward pathway and the audience's response data in the four cognitive dimensions as the feedback pathway, a closed-loop system is constructed. The output of the creative strategy network is the creative semantic attribute component, and the audience's response data in the four cognitive dimensions is the feedback quantity of the closed-loop system. Plot the Nyquist curve of the closed-loop system on the complex plane, calculate the minimum value of the Euclidean distance from the complex coordinates corresponding to all sampling frequency points of the Nyquist curve to the preset critical point of the Nyquist criterion as the Nyquist proximity, and use the Nyquist proximity as the system stability criterion value. When the system stability criterion value is lower than the preset stability margin threshold, it is determined that there is a coupled resonance effect in the closed-loop system; In the transfer function matrix, transfer functions with a damping ratio less than a preset critical damping value and a gain coefficient exceeding a preset gain threshold are selected. The creative semantic attribute components corresponding to the selected transfer functions are identified as the semantic attribute components of the resonance source.

[0011] Furthermore, suppression is applied to the creative semantic attribute generation sub-paths corresponding to the semantic attribute components of the resonance source, specifically including: Based on the system transfer function of the closed-loop system and the preset stability margin threshold, the inverse solution is the minimum gain attenuation required to restore the Nyquist proximity to above the preset stability margin threshold. The minimum gain attenuation is converted into the suppression gain value of the creative semantic attribute generation sub-path on the orthogonal dimension corresponding to the semantic attribute components of the resonance source; In the creative strategy network, a suppression gain value is applied to the creative semantic attribute generation sub-path on the orthogonal dimension corresponding to the semantic attribute component of the resonance source, which limits the maximum intensity of the creative strategy network activating the corresponding orthogonal dimension in the subsequent generation process.

[0012] Furthermore, when acquiring audience response data across multiple cognitive dimensions after the initial ad copy is placed, this also includes: Obtain the core advertising metrics set by the business partner. These core metrics include conversion rate and return on investment. The advertising conversion process is broken down into the exposure stage, click stage, page dwell stage, add-to-cart stage, and payment stage. Assign exposure completion rate to the exposure phase, click pass rate to the click phase, effective dwell time to the page dwell phase, add-to-cart conversion rate to the add-to-cart phase, and payment completion rate to the payment phase. Calculate the trend deviation between the core advertising metrics and each micro-conversion rate metric over time, and identify micro-conversion rate metrics whose trend deviation exceeds a preset deviation threshold as micro-conversion bottleneck metrics. Based on the micro-conversion bottleneck index, the audience cognitive decay index is extracted from the audience's response data across multiple cognitive dimensions. The audience cognitive decay index characterizes the rate at which the audience's response intensity to a specific creative semantic attribute decays with the number of repeated exposures.

[0013] Furthermore, it also includes: The creative semantic attribute generation sub-path with applied suppression gain value is continuously monitored, and the creative semantic attribute generation sub-path with applied suppression gain value corresponds to the semantic attribute component of the resonance source; the Nyquist proximity of the closed-loop system is continuously calculated. When the Nyquist proximity recovers to above the preset recovery threshold, it is determined that the coupling resonance effect between the semantic attribute component of the resonance source and the audience's response data in the four cognitive dimensions has subsided. Remove the suppression gain value for the creative semantic attribute generation sub-path and restore the creative policy network's full generation capability in the corresponding orthogonal dimension.

[0014] Furthermore, before identifying creative semantic attribute components that exhibit coupling and resonance effects with audience response data across multiple cognitive dimensions based on system stability criterion values, the process also includes: Determine whether the data volume of the first advertising copy has reached the preset data volume threshold; When the amount of data delivered reaches the preset data volume threshold, the calculation of the system stability criterion value is triggered.

[0015] This invention provides an advertising content generation system based on natural language processing, comprising: The Creative Strategy Network Module is used to construct a Creative Semantic Orthogonal Decoupling Space, which contains multiple orthogonal dimensions, each corresponding to an independent creative semantic attribute. The module maps the contextual signals of the advertising scenario to creative semantic attribute components in the Creative Semantic Orthogonal Decoupling Space, generating the first ad copy for delivery. The coupling relationship model construction module is used to obtain the audience's response data in multiple cognitive dimensions after the first advertising copy is delivered. Based on the relationship between the creative semantic attribute components and the audience's response data in multiple cognitive dimensions, a coupling relationship model between the creative semantic attribute components and the audience's response data in multiple cognitive dimensions is established. The resonance source identification module is used to identify creative semantic attribute components that have a coupling resonance effect with the audience's response data in multiple cognitive dimensions in the coupling relationship model based on the system stability criterion value, and to use them as resonance source semantic attribute components. The resonance suppression module is used to suppress the creative semantic attribute generation sub-path corresponding to the semantic attribute component of the resonance source in response to a new ad delivery request, and the creative strategy network module generates a second ad copy for delivery.

[0016] The beneficial effects of this invention compared to existing technologies are as follows: Existing technologies model the relationship between advertising creativity and audience cognition as a unidirectional decay process, triggering adjustments to the delivery strategy or replacing the entire copy solely by monitoring whether the cognitive response data exceeds a preset threshold. This approach fails to identify the coupling resonance effect between specific creative semantic attributes generated by the natural language processing model and the audience's deep psychological schema, leading to divergent oscillations between the output generated by the creative strategy network and the audience's cognitive state, preventing the system from converging to a stable matching state. This invention establishes a coupling relationship model between creative semantic attribute components and audience response data across multiple cognitive dimensions. Based on system stability criteria, it identifies resonance source semantic attribute components that exhibit coupling resonance effects with audience cognitive dimensions. Then, it applies targeted suppression to the corresponding generation sub-paths in the creative strategy network before iteratively generating advertising copy, eliminating the resonance instability state of the creative-audience cognitive coupling system at its root, restoring the system to a stable convergence state. While maintaining the other semantic dimension outputs of the copy generated by the creative strategy network unchanged, it only targets and suppresses the specific creative semantic dimensions that cause system instability, thereby improving the long-term output stability of the copy generated by the creative strategy network.

[0017] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in this application.

[0018] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0019] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart illustrating the overall technical solution of the advertising content generation method based on natural language processing in this embodiment of the invention. Figure 2 This is an expanded diagram of the coupling relationship model establishment and stability determination process in an embodiment of the present invention. Detailed Implementation

[0020] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0021] refer to Figure 1 and Figure 2 This invention provides an embodiment of an advertising content generation method based on natural language processing, comprising: Construct a creative semantic orthogonal decoupling space, which contains multiple orthogonal dimensions, each corresponding to an independent creative semantic attribute; deploy a creative strategy network to map the context signals of the advertising scenario to creative semantic attribute components in the creative semantic orthogonal decoupling space, and generate the first ad copy for delivery; After the first ad copy is delivered, the audience's response data in multiple cognitive dimensions is obtained. Based on the relationship between the creative semantic attribute components and the audience's response data in multiple cognitive dimensions, a coupling relationship model between the creative semantic attribute components and the audience's response data in multiple cognitive dimensions is established. In the coupling relationship model, based on the system stability criterion, creative semantic attribute components that have a coupling resonance effect with the audience's response data in multiple cognitive dimensions are identified as the resonance source semantic attribute components. In response to a new ad delivery request, the creative semantic attribute generation sub-path corresponding to the semantic attribute component of the resonance source is suppressed, and the creative strategy network generates a second ad copy for delivery.

[0022] In this embodiment, constructing a creative semantic orthogonal decoupling space refers to building a vector space for decoupling the creative features of advertising copy. The construction method of this creative semantic orthogonal decoupling space is as follows: First, multiple orthogonal dimensions are determined, each corresponding to an independent creative semantic attribute, including appeal type, emotional tone, language style, rhetorical devices, and narrative structure; then, an orthogonal basis vector set is constructed, where any two basis vectors in the orthogonal basis vector set are mutually orthogonal, and each basis vector corresponds to an orthogonal dimension. This creative semantic orthogonal decoupling space is spanned by the orthogonal basis vector set.

[0023] In this embodiment, the creative semantic orthogonal decoupling space contains multiple orthogonal dimensions, each corresponding to an independent creative semantic attribute. Changes in the coordinate values ​​of one orthogonal dimension do not affect the coordinate values ​​of other orthogonal dimensions, thus ensuring that each creative semantic attribute is independent of the others in the vector space.

[0024] In this embodiment, deploying the creative strategy network refers to deploying a neural network model that converts the contextual signals of the advertising delivery scenario into creative semantic attribute components. The creative strategy network adopts an encoder-decoder language model based on the Transformer architecture. The input of the creative strategy network is the contextual signals of the advertising delivery scenario, and the output is the creative semantic attribute components in the creative semantic orthogonal decoupling space.

[0025] In this embodiment, the contextual signals of the advertising placement scenario refer to the structured data describing the current advertising placement scenario, including the placement channel type, target audience profile tags, advertising product category, placement time period, and bidding keywords.

[0026] In this embodiment, mapping the contextual signals of the advertising scenario to creative semantic attribute components in the creative semantic orthogonal decoupling space means inputting the contextual signals of the advertising scenario into the creative strategy network, which then generates a vector in the creative semantic orthogonal decoupling space. This vector represents the creative semantic attribute component. The projection value of the creative semantic attribute component in each orthogonal dimension represents the activation intensity of the corresponding creative semantic attribute.

[0027] In this embodiment, generating the first advertising copy for delivery means that the creative strategy network decodes and generates natural language text based on the creative semantic attribute components, which serves as the first advertising copy, and sends the first advertising copy to the target audience through the advertising delivery platform.

[0028] In this embodiment, the creative strategy network generates natural language text based on creative semantic attribute components. Specifically, the decoder of the creative strategy network receives creative semantic attribute components as conditional input, concatenates the creative semantic attribute components with the hidden layer representations of each layer of the decoder, predicts the probability distribution of the next word on a word-by-word basis, samples words from the probability distribution, and uses the sampled word sequence as the generated advertising copy. When the creative strategy network generates the second advertising copy, it uses the same decoding method as when generating the first advertising copy.

[0029] In this embodiment, acquiring the audience's response data across multiple cognitive dimensions after the first advertising copy is delivered refers to collecting user behavior data generated by the audience within a preset time window after the first advertising copy is delivered. This user behavior data includes page dwell time, scroll depth, click hotspot distribution, and number of replays. Feature extraction is performed on the user behavior data to obtain user behavior feature vectors. These vectors are then input into a preset behavior-cognition mapping model to calculate the response value corresponding to each cognitive dimension. The response values ​​for each cognitive dimension are combined to form the audience's response data across multiple cognitive dimensions. The behavior-cognition mapping model is a multi-output regression model based on a gradient boosting decision tree.

[0030] In this embodiment, the cognitive dimension refers to the measurement dimension used to quantify the characteristics of the audience's mental schema. The cognitive dimension includes cognitive novelty seeking degree, cognitive fatigue accumulation degree, social proof sensitivity, and anchoring effect bias degree. Cognitive novelty seeking degree characterizes the audience's preference for novel creative stimuli; cognitive fatigue accumulation degree characterizes the audience's accumulated fatigue with repeatedly exposed creative ideas; social proof sensitivity characterizes the audience's sensitivity to the influence of social group evaluation; and anchoring effect bias degree characterizes the degree of deviation of the audience from the anchoring effect of initial information.

[0031] In this embodiment, the relationship between the creative semantic attribute components and the audience's response data across multiple cognitive dimensions refers to the input-output mapping relationship between the projection values ​​of the creative semantic attribute components on each orthogonal dimension and the audience's response values ​​on each cognitive dimension. This relationship reflects how changes in the creative semantic attributes cause changes in the audience's cognitive state.

[0032] In this embodiment, based on the relationship between creative semantic attribute components and audience response data across multiple cognitive dimensions, a coupling relationship model between creative semantic attribute components and audience response data across multiple cognitive dimensions is established. Specifically, this involves: acquiring market feedback sequence data generated after the historical advertising copy was placed, where historical advertising copy refers to advertising copy placed before the first advertising copy was placed; forming input-output sequence pairs by combining the creative semantic attribute components of the historical advertising copy with the audience response data across multiple cognitive dimensions within the corresponding placement period; using a prediction error method based on gradient descent optimization to systematically identify the input-output sequence pairs, estimating the gain coefficient, natural frequency, and damping ratio of the transfer function between each pair of creative semantic attribute components and the audience response data across multiple cognitive dimensions, where the transfer function adopts a second-order oscillatory element form; and constructing a transfer function matrix from the transfer functions, where the element in the i-th row and j-th column of the transfer function matrix is ​​the transfer function between the i-th creative semantic attribute component and the j-th cognitive dimension, and the transfer function matrix serves as the coupling relationship model between creative semantic attribute components and audience response data across multiple cognitive dimensions.

[0033] In this embodiment, the system stability criterion value refers to a quantitative judgment criterion based on the Nyquist stability criterion. Specifically, it is calculated that the minimum value among the Euclidean distances from the complex coordinates corresponding to all sampling frequency points of the Nyquist curve of the closed-loop system to the preset critical point of the Nyquist criterion is used as the Nyquist proximity, and the Nyquist proximity is used as the system stability criterion value. When the system stability criterion value is lower than the preset stability margin threshold, the system stability criterion value indicates that the system tends to become unstable.

[0034] In this embodiment, the coupling resonance effect refers to the system instability phenomenon in which the projected value on a specific orthogonal dimension of the creative semantic attribute component and the response value of a specific cognitive dimension in the audience's response data across multiple cognitive dimensions form a positive feedback loop through the transfer function in the coupling relationship model, resulting in divergent oscillations between the generated output of the creative strategy network and the audience's cognitive state.

[0035] In this embodiment, based on the system stability criterion, creative semantic attribute components that exhibit coupling resonance effects with audience response data across multiple cognitive dimensions are identified as resonance source semantic attribute components. Specifically, a closed-loop system is constructed by using a creative strategy network as the forward path and audience response data across multiple cognitive dimensions as the feedback path. The output of the creative strategy network is the creative semantic attribute component, and the audience response data across multiple cognitive dimensions is the feedback quantity of the closed-loop system. The Nyquist curve of the closed-loop system is plotted on the complex plane, and the Nyquist proximity is calculated. When the system stability criterion is lower than a preset stability margin threshold, a coupling resonance effect is determined to exist in the closed-loop system. In the transfer function matrix, transfer functions with damping ratios less than a preset critical damping value and gain coefficients exceeding a preset gain threshold are selected, and the creative semantic attribute components corresponding to the selected transfer functions are identified as resonance source semantic attribute components.

[0036] In this embodiment, a new ad delivery request refers to the next ad delivery request initiated by the business party after the first ad copy has been delivered and the semantic attribute components of the resonance source have been identified. The context signal of the ad delivery scenario in the new ad delivery request is the same as the context signal of the ad delivery scenario in the first ad copy, and belongs to the subsequent delivery of the same ad delivery plan.

[0037] In this embodiment, suppression is applied to the creative semantic attribute generation sub-path corresponding to the semantic attribute component of the resonance source, and the creative strategy network generates a second advertising copy for delivery. Specifically, based on the system transfer function of the closed-loop system and a preset stability margin threshold, the minimum gain attenuation required to restore the Nyquist proximity to above the preset stability margin threshold is solved using the Newton iteration method. The minimum gain attenuation is used as the suppression gain value for the creative semantic attribute generation sub-path on the orthogonal dimension corresponding to the semantic attribute component of the resonance source. In the creative strategy network, the suppression gain value is applied to the creative semantic attribute generation sub-path on the orthogonal dimension corresponding to the semantic attribute component of the resonance source, limiting the maximum intensity of activation of the corresponding orthogonal dimension by the creative strategy network in subsequent generation processes. The creative strategy network with the suppression gain value applied generates a new creative semantic attribute component in the creative semantic orthogonal decoupling space, and the second advertising copy is generated based on the new creative semantic attribute component for delivery. The creative semantic attribute generation sub-path refers to the network computation path responsible for generating the projection value on a specific orthogonal dimension during the generation of the creative semantic attribute component by the creative strategy network. The suppression gain is applied by multiplying the logits value of the semantic attribute component of the resonant source on the orthogonal dimension by the suppression gain in the output layer of the creative policy network decoder, and then performing softmax normalization.

[0038] Furthermore, a creative semantic orthogonal decoupling space is constructed, specifically including: Multiple orthogonal dimensions are defined, each corresponding to an independent creative semantic attribute. Creative semantic attributes include appeal type, emotional tone, language style, rhetorical devices, and narrative structure. Construct an orthogonal basis vector set, in which any two basis vectors are orthogonal to each other, and each basis vector corresponds to an orthogonal dimension; The creative semantic attribute components are generated by the creative policy network in the creative semantic orthogonal decoupled space spanned by the orthogonal basis vectors. The projection value of the creative semantic attribute component on each orthogonal dimension represents the activation intensity of the corresponding creative semantic attribute.

[0039] In this embodiment, the creative semantic attributes include appeal type, emotional tone, language style, rhetorical devices, and narrative structure. Appeal type represents the core persuasive logic of the advertising copy, including rational appeals, emotional appeals, and fear appeals. Emotional tone represents the overall emotional color of the advertising copy, including positive, neutral, and negative tones. Language style represents the word choice and sentence structure characteristics of the advertising copy, including formal, colloquial, and humorous styles. Rhetorical devices represent the rhetorical techniques used in the advertising copy, including metaphor, parallelism, and rhetorical questions. Narrative structure represents the narrative organization method of the advertising copy, including problem-solving structure, testimony structure, and suspense structure.

[0040] In this embodiment, an orthogonal basis vector set is constructed, where any two basis vectors in the orthogonal basis vector set are mutually orthogonal. Each basis vector corresponds to an orthogonal dimension. Specifically, a basis vector is assigned to each orthogonal dimension of the creative semantic attribute, and all basis vectors constitute the orthogonal basis vector set. The inner product of the a-th basis vector and the b-th basis vector in the orthogonal basis vector set is zero, where a is not equal to b. The orthogonal basis vector set is constructed by performing Gram-Schmidt orthogonalization on a randomly initialized vector set to ensure that any two basis vectors are mutually orthogonal.

[0041] In this embodiment, the creative semantic orthogonal decoupling space spanned by the orthogonal basis vector set refers to the vector space generated with the orthogonal basis vector set as the basis. Any vector in the creative semantic orthogonal decoupling space can be represented as a linear combination of the basis vectors in the orthogonal basis vector set, where the combination coefficients are the coordinate values ​​of the vector in each orthogonal dimension. Since any two basis vectors in the orthogonal basis vector set are mutually orthogonal, a change in the coordinate value in one orthogonal dimension does not affect the coordinate values ​​in other orthogonal dimensions.

[0042] In this embodiment, the creative semantic attribute components are generated by the creative strategy network in a creative semantic orthogonal decoupling space spanned by a set of orthogonal basis vectors. Specifically, the creative strategy network receives the context signal of the advertising scenario as input and outputs a vector as a creative semantic attribute component in the creative semantic orthogonal decoupling space. The process of generating creative semantic attribute components by the creative strategy network is constrained by the set of orthogonal basis vectors. The output layer weight matrix of the creative strategy network is multiplied element-wise with the set of orthogonal basis vectors, so that the output vector is confined within the creative semantic orthogonal decoupling space.

[0043] In this embodiment, the projection value of the creative semantic attribute component on each orthogonal dimension represents the activation intensity of the corresponding creative semantic attribute. Specifically, the inner product operation is performed between the creative semantic attribute component and the basis vector corresponding to each orthogonal dimension to obtain the projection value of the creative semantic attribute component on that orthogonal dimension. The larger the absolute value of the projection value, the higher the activation intensity of the corresponding creative semantic attribute; a positive projection value indicates that the creative semantic attribute is activated with a positive tendency, and a negative projection value indicates that the creative semantic attribute is activated with a negative tendency. For example, when the projection value on the emotional tone dimension is positive and the absolute value is large, it indicates that the emotional tone of the advertising copy is positive and the activation intensity is high.

[0044] Furthermore, after the initial ad copy was placed, the audience's response data across multiple cognitive dimensions was obtained, specifically including: The audience cognitive state space is constructed, which includes four cognitive dimensions: cognitive novelty seeking degree, cognitive fatigue accumulation degree, social identity sensitivity degree, and anchoring effect deviation degree. These four cognitive dimensions together constitute a four-dimensional representation of the audience's psychological schema characteristics. After the first ad copy is placed, collect user behavior data generated by the audience within a preset time window. User behavior data includes page dwell time, scroll depth, hotspot distribution of clicks, and number of times the ad was viewed. Feature extraction is performed on user behavior data to obtain user behavior feature vectors; the user behavior feature vectors are input into a preset behavior-cognition mapping model to calculate the response value corresponding to each cognitive dimension; the response values ​​corresponding to the four cognitive dimensions are combined to form the audience's response data on multiple cognitive dimensions.

[0045] In this embodiment, constructing the audience cognitive state space refers to constructing a vector space for quantifying the characteristics of the audience's mental schema. The audience cognitive state space uses cognitive novelty seeking degree, cognitive fatigue accumulation degree, social identity sensitivity, and anchoring effect bias degree as coordinate axes. Each vector in the audience cognitive state space represents the audience's cognitive state in four cognitive dimensions.

[0046] In this embodiment, cognitive novelty seeking degree characterizes the audience's preference for novel creative stimuli. The higher the response value of cognitive novelty seeking degree, the stronger the audience's willingness to accept the semantic attributes of creatives that they have never encountered before. Cognitive fatigue accumulation degree characterizes the degree of fatigue accumulation of the audience to repeatedly exposed creatives. The higher the response value of cognitive fatigue accumulation degree, the stronger the audience's resistance to the semantic attributes of creatives that have been exposed many times. Social identity sensitivity characterizes the audience's sensitivity to the influence of social group evaluation. The higher the response value of social identity sensitivity, the greater the influence of others' evaluation information on the audience when they encounter advertising copy. Anchoring effect bias degree characterizes the degree of deviation of the audience from the initial information anchoring effect. The higher the response value of anchoring effect bias degree, the stronger the audience's fixation on the semantic attributes of creatives they encounter for the first time.

[0047] In this embodiment, the four cognitive dimensions together constitute a four-dimensional representation of the audience's psychological schema features. These four cognitive dimensions—the degree of cognitive novelty seeking, the degree of cognitive fatigue accumulation, the degree of social identity sensitivity, and the degree of anchoring effect deviation—describe the schema features of the audience when they process advertising creative ideas from different perspectives. The response values ​​of the four cognitive dimensions are combined to form a complete four-dimensional vector, which serves as a comprehensive quantitative representation of the audience's psychological schema features.

[0048] In this embodiment, the preset time window refers to a continuous time interval defined by a preset duration, starting from the moment the first advertisement copy is placed. The preset duration is set according to the advertising channel and the category of the advertised product, and its value ranges from 24 hours to 72 hours. The preset time window is used to limit the time range for collecting user behavior data.

[0049] In this embodiment, collecting user behavior data generated by the audience within a preset time window after the first advertising copy is placed refers to obtaining user behavior data generated by the audience during their interaction with the first advertising copy within a preset time window through the data collection interface of the advertising platform.

[0050] In this embodiment, the page dwell time is the length of time the audience stays on the landing page of the first advertisement copy, measured in seconds; the scroll depth is the percentage of the maximum scroll distance of the audience on the landing page of the first advertisement copy to the total height of the page; the click hot zone distribution is the set of click coordinates of the audience on the landing page of the first advertisement copy; and the number of times the audience revisits the first advertisement copy landing page is closed and then reopened.

[0051] In this embodiment, the user behavior feature vector is obtained by extracting features from user behavior data. This means taking the natural logarithm of the page dwell time as the dwell time feature, taking the maximum scroll percentage as the scroll depth feature, counting the click frequency within each preset hot zone grid as the click hot zone feature, and directly taking the number of times the user revisited as the revisit number feature. The dwell time feature, scroll depth feature, click hot zone feature, and revisit number feature are concatenated into a fixed-length numerical vector, which is the user behavior feature vector.

[0052] In this embodiment, the preset behavior-cognition mapping model refers to a pre-trained machine learning model used to map user behavior feature vectors to response values ​​in four cognitive dimensions. The behavior-cognition mapping model adopts a multi-output regression model based on gradient boosting decision trees. The training method of the behavior-cognition mapping model is as follows: collect user behavior data after historical advertising copy placement and extract features to obtain training behavior feature vectors; by embedding a survey questionnaire component on the advertising landing page, the questionnaire is triggered after the audience completes the browsing behavior, and the audience's user behavior data is associated with the audience's questionnaire rating results through session identifiers to obtain the audience's ratings on cognitive novelty seeking degree, cognitive fatigue accumulation degree, social identity sensitivity, and anchoring effect bias degree as training labels; using the training behavior feature vectors as input and the training labels as output, the multi-output gradient boosting decision tree model is trained using the mean squared error loss function and gradient descent optimization algorithm to obtain the behavior-cognition mapping model.

[0053] In this embodiment, inputting the user behavior feature vector into a preset behavior-cognition mapping model to calculate the response value corresponding to each cognitive dimension means that the user behavior feature vector is used as the input of the behavior-cognition mapping model, and the behavior-cognition mapping model outputs four values, which correspond to the response values ​​of cognitive novelty seeking degree, cognitive fatigue accumulation degree, social identity sensitivity degree, and anchoring effect bias degree, respectively.

[0054] In this embodiment, combining the response values ​​corresponding to the four cognitive dimensions to form the audience's response data across multiple cognitive dimensions means concatenating the response values ​​of cognitive novelty seeking degree, cognitive fatigue accumulation degree, social identity sensitivity degree, and anchoring effect deviation degree in a fixed order into a four-dimensional vector, which represents the audience's response data across multiple cognitive dimensions.

[0055] Furthermore, based on the relationship between creative semantic attribute components and audience response data across multiple cognitive dimensions, a coupling relationship model between creative semantic attribute components and audience response data across multiple cognitive dimensions is established, specifically including: Obtain market feedback sequence data generated after the historical ad copy was launched. The historical ad copy refers to the ad copy that was launched before the first ad copy was launched. Combine the creative semantic attribute components of the historical ad copy with the audience's response data in four cognitive dimensions during the corresponding campaign period to form input-output sequence pairs. The input-output sequence pairs are systematically identified, and the gain coefficient, natural frequency, and damping ratio of the transfer function between each pair of creative semantic attribute components and the audience's response data in one of the four cognitive dimensions are estimated. The transfer function adopts the form of a second-order oscillatory element. The transfer function matrix is ​​composed of transfer functions. The element in the i-th row and j-th column of the transfer function matrix is ​​the transfer function between the i-th creative semantic attribute component and the j-th cognitive dimension. The transfer function matrix serves as a coupling relationship model between the creative semantic attribute component and the audience's response data in the four cognitive dimensions.

[0056] In this embodiment, acquiring market feedback sequence data generated after the historical ad copy placement refers to extracting the creative semantic attribute components of multiple historical ad copy placed before the first ad copy placement from the historical placement logs of the ad placement platform, as well as the audience's response data in four cognitive dimensions at each sampling time during the placement period of each historical ad copy. Each sampling time in the market feedback sequence data corresponds to a set of creative semantic attribute components and a set of audience response data in four cognitive dimensions.

[0057] In this embodiment, the campaign period refers to a continuous time interval from the moment an advertisement is first launched to the moment when the advertisement is no longer launched. The campaign period is divided into multiple sampling times according to a preset sampling interval, which is set based on the frequency of advertisement launches. The audience's response data across the four cognitive dimensions within a campaign period are arranged in chronological order of the sampling times, forming a time series of the audience's response data across the four cognitive dimensions within that campaign period.

[0058] In this embodiment, the creative semantic attribute components of historical advertising copy are combined with the audience's response data across four cognitive dimensions within the corresponding campaign period to form an input-output sequence pair. This means that for each historical advertising copy, the creative semantic attribute components of that copy are used as the input sequence, and the audience's response data across the four cognitive dimensions at each sampling time within the campaign period are used as the output sequence, forming an input-output sequence pair. The input sequence in the input-output sequence pair is a creative semantic attribute component, and the output sequence is a time series, with the length of the output sequence equal to the number of sampling times within the campaign period.

[0059] In this embodiment, system identification is performed on the input-output sequence pairs. The gain coefficient, natural frequency, and damping ratio of the transfer function between each pair of creative semantic attribute components and the audience's response data in one of the four cognitive dimensions are estimated. The transfer function adopts a second-order oscillatory form. Specifically, the projection values ​​of each orthogonal dimension in the creative semantic attribute components are used as the input signal for system identification, and the time series of the audience's response data in one of the four cognitive dimensions is used as the output signal for system identification. The structure of the transfer function is preset to a second-order oscillatory form, which is the product of the gain coefficient and the second-order oscillation term. The second-order oscillation term is determined by the natural frequency and the damping ratio. A prediction error method based on gradient descent optimization is used to iteratively update the estimated values ​​of the gain coefficient, natural frequency, and damping ratio with the goal of minimizing the mean square error between the predicted and actual values ​​of the output signal, until the mean square error converges to below a preset convergence threshold. During the iterative update process, non-negativity constraints are applied to the gain coefficient, natural frequency, and damping ratio. The projective gradient descent method is used to project the parameter values ​​after each iteration onto the feasible region that satisfies the constraints, obtaining the final estimates of the gain coefficient, natural frequency, and damping ratio of the transfer function. This process is performed separately between the projected values ​​on each orthogonal dimension of each creative semantic attribute component and the response data of each cognitive dimension, resulting in a total of transfer functions equal to the number of dimensions of the creative semantic orthogonal decoupling space multiplied by the number of cognitive dimensions.

[0060] In this embodiment, a transfer function matrix is ​​constructed from transfer functions. The element in the i-th row and j-th column of the transfer function matrix is ​​the transfer function between the i-th creative semantic attribute component and the j-th cognitive dimension. The transfer function matrix serves as a coupling relationship model between the creative semantic attribute component and the audience's response data in the four cognitive dimensions. Specifically, the number of dimensions of the creative semantic orthogonal decoupling space is denoted as N, and the number of dimensions of the audience's cognitive state space is denoted as M. An N-row, M-column matrix is ​​constructed as the transfer function matrix. The transfer function between the i-th creative semantic attribute component and the j-th cognitive dimension is filled into the i-th row and j-th column of the transfer function matrix. Each row of the transfer function matrix corresponds to the projection value of an orthogonal dimension in a creative semantic attribute component, and each column corresponds to a cognitive dimension. Each element in the matrix is ​​a transfer function, which fully describes the dynamic response relationship where the projection value in the orthogonal dimension is used as input, causing the response value in the cognitive dimension to be used as output. The transfer function matrix serves as a coupling model between the creative semantic attribute components and the audience's response data across four cognitive dimensions. The input is a vector composed of the projection values ​​of each orthogonal dimension of the creative semantic attribute components, and the output is a vector composed of the predicted values ​​of the audience's response values ​​across the four cognitive dimensions.

[0061] Furthermore, in the coupling relationship model, based on the system stability criterion, creative semantic attribute components that exhibit coupling resonance effects with the audience's response data across multiple cognitive dimensions are identified. These components specifically include: Using the creative strategy network as the forward pathway and the audience's response data in the four cognitive dimensions as the feedback pathway, a closed-loop system is constructed. The output of the creative strategy network is the creative semantic attribute component, and the audience's response data in the four cognitive dimensions is the feedback quantity of the closed-loop system. Plot the Nyquist curve of the closed-loop system on the complex plane, calculate the minimum value of the Euclidean distance from the complex coordinates corresponding to all sampling frequency points of the Nyquist curve to the preset critical point of the Nyquist criterion as the Nyquist proximity, and use the Nyquist proximity as the system stability criterion value. When the system stability criterion value is lower than the preset stability margin threshold, it is determined that there is a coupled resonance effect in the closed-loop system; In the transfer function matrix, transfer functions with a damping ratio less than a preset critical damping value and a gain coefficient exceeding a preset gain threshold are selected. The creative semantic attribute components corresponding to the selected transfer functions are identified as the semantic attribute components of the resonance source.

[0062] In this embodiment, a creative strategy network is used as the forward path, and the audience's response data in four cognitive dimensions is used as the feedback path to construct a closed-loop system. Specifically, the creative semantic attribute components output by the creative strategy network are used as the output signal of the forward path. These components are input into a coupling relationship model, which outputs predicted values ​​of the audience's response data in the four cognitive dimensions. These predicted values ​​are then used as the output signal of the feedback path. The feedback path's output signal is compared with a preset reference input signal to obtain an error signal, which is then input into the creative strategy network to form a closed loop. The reference input signal for the closed-loop system is a zero vector, representing the desired stable level of the audience's cognitive state. The output of the creative strategy network is the creative semantic attribute component, and the audience's response data in the four cognitive dimensions serves as the feedback quantity for the closed-loop system. This feedback quantity is used to compare with the reference input signal to form closed-loop control. This closed-loop system is a virtual analysis system built based on the coupling relationship model. Its function is to provide an analytical framework and computational basis for identifying the resonance source semantic attribute components that cause system instability and calculating the gain attenuation required to restore system stability. In the actual operation of advertising content generation and delivery, the input signal of the creative strategy network is the context signal of the advertising delivery scenario. There is no real-time negative feedback control loop between the creative strategy network and the audience's cognitive response. Modification of the creative strategy network is achieved by applying the suppression gain value calculated based on the virtual analysis system to the creative semantic attribute generation sub-path, which is a parameter adjustment guided by the analysis results.

[0063] In this embodiment, the complex plane refers to a two-dimensional planar coordinate system with the horizontal axis as the real axis and the vertical axis as the imaginary axis. The complex plane is used to plot the Nyquist curve. Each point on the complex plane corresponds to a complex number, which is determined by its real and imaginary coordinates.

[0064] In this embodiment, the Nyquist curve of the closed-loop system is plotted on the complex plane. Specifically, the complex frequency response values ​​calculated by the open-loop transfer function of the closed-loop system at different frequency values ​​are mapped onto the complex plane. The complex points corresponding to each frequency value are connected sequentially in ascending order from zero to positive infinity to form a continuous curve as the Nyquist curve. The open-loop transfer function is obtained by multiplying the transfer function of the forward path and the transfer function of the feedback path. The transfer function of the forward path is the equivalent transfer function of the creative strategy network. The equivalent transfer function is obtained by linearly combining the transfer functions in the row vectors corresponding to the semantic attribute components of the resonance source in the transfer function matrix identified by the system in this embodiment. The coefficients of the linear combination are the normalized weights of the projection values ​​of the semantic attribute components of the resonance source in each orthogonal dimension. The transfer function of the feedback path is obtained by linearly combining the transfer functions in the row vectors corresponding to the semantic attribute components of the resonance source in the transfer function matrix of the coupling relationship model with the weight coefficients of each cognitive dimension of the audience's response data in the four cognitive dimensions.

[0065] In this embodiment, the object of Nyquist analysis is the equivalent linear system described by the transfer function matrix identified by the system. The equivalent transfer function of the forward path is obtained by linearly combining the transfer functions in the row vectors corresponding to the semantic attribute components of the resonant source in the transfer function matrix, and the coefficients of the linear combination are the normalized weights of the projection values ​​of the semantic attribute components of the resonant source in each orthogonal dimension. The equivalent transfer function of the feedback path is obtained by linearly combining the transfer functions in the row vectors corresponding to the semantic attribute components of the resonant source in the transfer function matrix with the weight coefficients of each cognitive dimension of the audience's response data in the four cognitive dimensions. The open-loop transfer function is the product of the equivalent transfer function of the forward path and the equivalent transfer function of the feedback path. The plotting of the Nyquist curve and the calculation of the Nyquist proximity are both based on the open-loop transfer function of this equivalent linear system. The above-mentioned construction method of the equivalent linear system allows Nyquist analysis to complete the stability determination by utilizing the equivalent dynamic relationship between the creative semantic attribute components identified by the system and the audience's cognitive response, without needing to obtain the analytical transfer function of the creative strategy network itself.

[0066] In this embodiment, all sampling frequency points refer to multiple frequency values ​​selected from zero to a preset maximum frequency range at preset frequency intervals when plotting the Nyquist curve. The preset frequency interval and preset maximum frequency are determined based on the sampling frequency of the audience's response data in the four cognitive dimensions. The preset maximum frequency is the Nyquist frequency of the sampling frequency of the audience's response data in the four cognitive dimensions, and the preset frequency interval is the Nyquist frequency divided by the preset number of sampling points.

[0067] In this embodiment, the complex coordinates corresponding to all sampling frequency points of the Nyquist curve refer to the coordinates of a point obtained by mapping the complex frequency response value of the open-loop transfer function at that sampling frequency point to the complex plane for each sampling frequency point. The complex coordinates consist of real and imaginary coordinates, where the real coordinates are equal to the real part of the complex frequency response value and the imaginary coordinates are equal to the imaginary part of the complex frequency response value.

[0068] In this embodiment, the Nyquist criterion preset critical point refers to the reference point used in the Nyquist stability criterion to determine the stability of the closed-loop system. The coordinates of the Nyquist criterion preset critical point on the complex plane are -1 plus zero multiplied by the imaginary unit.

[0069] In this embodiment, the minimum Euclidean distance from the complex coordinates corresponding to all sampling frequency points of the Nyquist curve to the preset critical point of the Nyquist criterion is calculated. Specifically, for each complex coordinate corresponding to a sampling frequency point, the square of the difference between the real part of the complex coordinate and the real part of the preset critical point of the Nyquist criterion is added to the square of the difference between the imaginary part of the complex coordinate and the imaginary part of the preset critical point of the Nyquist criterion. The square root of this sum is taken to obtain the Euclidean distance from the complex coordinate corresponding to that sampling frequency point to the preset critical point of the Nyquist criterion. The minimum value among all Euclidean distances corresponding to all sampling frequency points is taken as the Nyquist proximity score, and the Nyquist proximity score is taken as the system stability criterion value.

[0070] In this embodiment, the preset stability margin threshold refers to a pre-set critical value for Nyquist proximity used to determine whether the closed-loop system tends towards instability. The preset stability margin threshold is set according to the tolerance of the advertising delivery business for system stability.

[0071] In this embodiment, determining that there is a coupling resonance effect in the closed-loop system means that when the Nyquist proximity is lower than the preset stability margin threshold, the Nyquist curve approaches the preset critical point of the Nyquist criterion, the relative stability margin of the closed-loop system is insufficient, and the generated output of the creative strategy network and the audience's cognitive state enter a positive feedback self-excitation state, and the system has a coupling resonance effect.

[0072] In this embodiment, within the transfer function matrix, the creative semantic attribute components corresponding to transfer functions with damping ratios less than a preset critical damping value and gain coefficients exceeding a preset gain threshold are identified as resonance source semantic attribute components. Specifically, this involves: traversing all transfer functions in the transfer function matrix and filtering out those with damping ratios less than the preset critical damping value; comparing the absolute values ​​of the gain coefficients of each selected transfer function to determine the row containing the transfer function with the largest absolute gain coefficient; and identifying the creative semantic attribute component corresponding to that row as the resonance source semantic attribute component. A damping ratio less than the preset critical damping value is the core criterion for oscillating resonance in a second-order system. A large absolute value of the gain coefficient indicates that the creative semantic attribute component has the greatest amplification effect on the audience's cognitive dimension response. The combined determination of these two factors can accurately pinpoint the primary driving source of the coupled resonance effect. The preset critical damping value is set to 0.707, and the preset gain threshold is set to 2.

[0073] Furthermore, suppression is applied to the creative semantic attribute generation sub-paths corresponding to the semantic attribute components of the resonance source, specifically including: Based on the system transfer function of the closed-loop system and the preset stability margin threshold, the inverse solution is the minimum gain attenuation required to restore the Nyquist proximity to above the preset stability margin threshold. The minimum gain attenuation is converted into the suppression gain value of the creative semantic attribute generation sub-path on the orthogonal dimension corresponding to the semantic attribute components of the resonance source; In the creative strategy network, a suppression gain value is applied to the creative semantic attribute generation sub-path on the orthogonal dimension corresponding to the semantic attribute component of the resonance source, which limits the maximum intensity of the creative strategy network activating the corresponding orthogonal dimension in the subsequent generation process.

[0074] In this embodiment, based on the system transfer function of the closed-loop system and a preset stability margin threshold, the minimum gain attenuation required to restore the Nyquist proximity to above the preset stability margin threshold is determined by inverse kinematics. Specifically, the Nyquist proximity is used as a monotonically decreasing function of the gain coefficient of the open-loop transfer function of the closed-loop system. Starting from the current gain coefficient value, the gain coefficient value is gradually decreased with a preset step size. After each decrease, the corresponding Nyquist proximity is recalculated until the Nyquist proximity is restored to a value greater than or equal to the preset stability margin threshold. The difference between the initial gain coefficient value and the restored gain coefficient value is taken as the minimum gain attenuation. The preset step size is one percent of the initial gain coefficient value.

[0075] In this embodiment, the minimum gain attenuation is converted into a suppression gain value for the creative semantic attribute generation sub-path on the orthogonal dimension corresponding to the semantic attribute components of the resonance source. Specifically, the minimum gain attenuation is divided by the initial gain coefficient value to obtain the gain attenuation ratio; the suppression gain value is obtained by subtracting the gain attenuation ratio from one. The suppression gain value is a value greater than zero and less than or equal to one. A suppression gain value of one indicates that no suppression is applied, and the closer the suppression gain value is to zero, the stronger the suppression.

[0076] In this embodiment, within the creative strategy network, a suppression gain value is applied to the creative semantic attribute generation sub-path on the orthogonal dimension corresponding to the resonant source semantic attribute component. This limits the maximum intensity of activation of the corresponding orthogonal dimension by the creative strategy network in subsequent generation processes. Specifically, in the output layer of the creative strategy network decoder, the position of the logits component corresponding to the orthogonal dimension of the resonant source semantic attribute component is determined. Each time the creative strategy network generates the second advertising copy, the logits component corresponding to that orthogonal dimension in the decoder output layer is multiplied by the suppression gain value. Then, the logits vector multiplied by the suppression gain value is input into the softmax function for normalization to obtain the vocabulary probability distribution. The natural language text of the second advertising copy is generated based on the vocabulary probability distribution. By applying the suppression gain value, the maximum intensity of activation of the corresponding orthogonal dimension by the creative strategy network in subsequent generation processes is limited to the suppression gain value multiple level when no suppression is applied.

[0077] In this embodiment, the maximum strength of the orthogonal dimension refers to the maximum absolute value that the projection value on the corresponding orthogonal dimension can achieve without suppression when the creative strategy network generates creative semantic attribute components. The maximum strength is determined by the numerical range of the logits component corresponding to that orthogonal dimension in the output layer of the creative strategy network decoder. After applying a suppression gain value, the maximum absolute value of the projection value on the corresponding orthogonal dimension is limited to a multiple of the suppression gain value when no suppression is applied.

[0078] Furthermore, when acquiring audience response data across multiple cognitive dimensions after the initial ad copy is placed, this also includes: Obtain the core advertising metrics set by the business partner. These core metrics include conversion rate and return on investment. The advertising conversion process is broken down into the exposure stage, click stage, page dwell stage, add-to-cart stage, and payment stage. Assign exposure completion rate to the exposure phase, click pass rate to the click phase, effective dwell time to the page dwell phase, add-to-cart conversion rate to the add-to-cart phase, and payment completion rate to the payment phase. Calculate the trend deviation between the core advertising metrics and each micro-conversion rate metric over time, and identify micro-conversion rate metrics whose trend deviation exceeds a preset deviation threshold as micro-conversion bottleneck metrics. Based on the micro-conversion bottleneck index, the audience cognitive decay index is extracted from the audience's response data across multiple cognitive dimensions. The audience cognitive decay index characterizes the rate at which the audience's response intensity to a specific creative semantic attribute decays with the number of repeated exposures.

[0079] In this embodiment, obtaining the core advertising metrics set by the business party refers to reading the core metric data pre-set by the business party to measure the effectiveness of advertising through the configuration interface of the advertising platform. These core advertising metrics are set by the business party based on advertising objectives and serve as a benchmark for evaluating advertising effectiveness.

[0080] In this embodiment, conversion rate refers to the ratio of the number of viewers who completed a payment to the number of times the advertising copy was displayed. Return on investment (ROI) is the total sales revenue generated by the advertising campaign divided by the total cost of the advertising campaign.

[0081] In this embodiment, the advertising conversion process is broken down into five stages: exposure, click, page dwell, add-to-cart, and payment. This means that the advertising conversion process is divided into five stages according to the complete behavioral chain from the audience's initial contact with the advertising copy to the completion of the payment. The exposure stage corresponds to the process where the advertising copy is displayed to the audience on the advertising channel; the click stage corresponds to the process where the audience clicks on the advertising copy and enters the landing page; the page dwell stage corresponds to the process where the audience browses the content on the landing page; the add-to-cart stage corresponds to the process where the audience adds the product to their shopping cart; and the payment stage corresponds to the process where the audience completes the order payment.

[0082] In this embodiment, an exposure completion rate metric is assigned to the exposure stage, which is the ratio of the actual number of times the ad copy is exposed to the planned number of exposures; a click-through rate metric is assigned to the click stage, which is the ratio of the number of times the ad copy is clicked to the number of times the ad copy is exposed; an effective dwell time metric is assigned to the page dwell stage, which is the ratio of the number of sessions where the page dwell time exceeds a preset effective dwell time threshold to the number of times the ad copy is clicked; an add-to-cart conversion rate metric is assigned to the add-to-cart stage, which is the ratio of the number of times the ad copy is added to the cart to the number of effective dwell times; and a payment completion rate metric is assigned to the payment stage, which is the ratio of the number of times the payment is made to the number of times the ad copy is added to the cart.

[0083] In this embodiment, the micro-conversion rate metric refers to the collective term for the exposure completion rate, click-through rate, effective dwell time rate, add-to-cart conversion rate, and payment completion rate. Each micro-conversion rate metric corresponds to a stage in the advertising conversion process and is used to measure the conversion efficiency of that stage.

[0084] In this embodiment, the trend deviation between the core advertising metrics and each micro-conversion rate metric over time is calculated as follows: The core advertising metrics and each micro-conversion rate metric are obtained at each sampling moment within a preset time window after the first advertising copy is placed, forming the time series of the core advertising metrics and the time series of each micro-conversion rate metric; the time series of the core advertising metrics is linearly fitted to obtain the trend slope of the core advertising metrics, and the time series of each micro-conversion rate metric is linearly fitted to obtain the trend slope of each micro-conversion rate metric; the absolute value of the difference between the trend slope of each micro-conversion rate metric and the trend slope of the core advertising metrics is calculated, and this absolute value is used as the trend deviation between the micro-conversion rate metric and the core advertising metrics over time.

[0085] In this embodiment, the preset deviation threshold refers to a pre-set critical value for determining the degree of trend deviation of a micro-conversion rate indicator as a bottleneck in conversion. The preset deviation threshold is determined based on historical experience data from advertising campaigns.

[0086] In this embodiment, micro-conversion rate indicators whose trend deviation exceeds a preset deviation threshold are identified as micro-conversion bottleneck indicators. This means comparing the trend deviation of each micro-conversion rate indicator with the preset deviation threshold sequentially, and marking micro-conversion rate indicators with trend deviations greater than the preset deviation threshold as micro-conversion bottleneck indicators. Micro-conversion bottleneck indicators characterize the stage in the advertising conversion process where the conversion efficiency is significantly lower than the overall trend, representing a weak link that causes the core indicators of advertising to fail to meet expectations.

[0087] In this embodiment, based on the micro-conversion bottleneck indicator, an audience cognitive decay indicator is extracted from the audience's response data across multiple cognitive dimensions. Specifically, this involves: determining the stage of the advertising conversion process corresponding to the micro-conversion bottleneck indicator; extracting the response data of the cognitive dimension corresponding to that stage from the time series of the audience's response data across the four cognitive dimensions, and using the response data of the cognitive dimension corresponding to the micro-conversion bottleneck indicator as the audience cognitive decay indicator. The audience cognitive decay indicator is the response data of one cognitive dimension among the audience's response data across the four cognitive dimensions that has a corresponding relationship with the micro-conversion bottleneck indicator.

[0088] In this embodiment, the audience cognitive decay index extracted from the audience's response data in multiple cognitive dimensions is used as the response data of one cognitive dimension among the audience's response data in four cognitive dimensions. It is then input into the establishment process of the coupling relationship model described in claim 4 to update the transfer function parameters of the corresponding cognitive dimension in the transfer function matrix.

[0089] Furthermore, it also includes: The creative semantic attribute generation sub-path with applied suppression gain value is continuously monitored, and the creative semantic attribute generation sub-path with applied suppression gain value corresponds to the semantic attribute component of the resonance source; the Nyquist proximity of the closed-loop system is continuously calculated. When the Nyquist proximity recovers to above the preset recovery threshold, it is determined that the coupling resonance effect between the semantic attribute component of the resonance source and the audience's response data in the four cognitive dimensions has subsided. Remove the suppression gain value for the creative semantic attribute generation sub-path and restore the creative policy network's full generation capability in the corresponding orthogonal dimension.

[0090] In this embodiment, the creative semantic attribute generation sub-path with the applied suppression gain value is continuously monitored. The creative semantic attribute generation sub-path with the applied suppression gain value corresponds to the semantic attribute component of the resonant source. The Nyquist proximity of the closed-loop system is continuously calculated. Specifically, after the second advertising copy is generated and launched, the audience's response data in four cognitive dimensions is collected every preset monitoring period. The creative semantic attribute components of the second advertising copy and the collected audience response data in four cognitive dimensions are input into the coupling relationship model to update the transfer function matrix. Based on the updated transfer function matrix, the open-loop transfer function of the closed-loop system is updated, the Nyquist curve is redrawn on the complex plane, and the minimum value of the Euclidean distance from the complex coordinates corresponding to all sampling frequency points of the Nyquist curve to the preset critical point of the Nyquist criterion is recalculated as the updated Nyquist proximity. The preset monitoring period is set according to the advertising frequency and the collection period of audience response data.

[0091] In this embodiment, when the Nyquist proximity recovers to above the preset recovery threshold, it is determined that the coupling resonance effect between the semantic attribute component of the resonance source and the audience's response data in the four cognitive dimensions has subsided. Specifically, the relationship between the Nyquist proximity calculated in each monitoring cycle and the preset recovery threshold is compared, and the preset recovery threshold is greater than the preset stability margin threshold. When the Nyquist proximity reaches a level greater than or equal to the preset recovery threshold, it indicates that the stability margin of the closed-loop system has recovered from the critical unstable state to the fully stable state, the positive feedback self-excitation relationship between the semantic attribute component of the resonance source and the audience's response data in the four cognitive dimensions has been released, and the coupling resonance effect has subsided.

[0092] In this embodiment, the suppression gain value of the creative semantic attribute generation sub-path is revoked to restore the creative strategy network's full generation capability in the corresponding orthogonal dimension. Specifically, after determining that the coupling resonance effect has subsided, the suppression gain value of the creative semantic attribute generation sub-path in the orthogonal dimension corresponding to the resonance source semantic attribute component is reset to one, thus removing the multiplication suppression operation on the logits component corresponding to that orthogonal dimension in the decoder output layer. After the suppression gain value is revoked, the creative strategy network is no longer subject to activation intensity restrictions on the creative semantic attribute generation sub-path corresponding to that orthogonal dimension when generating subsequent advertising copy.

[0093] In this embodiment, the complete generation capability of the creative strategy network in the corresponding orthogonal dimension means that, without applying a suppression gain value, the logits component in the corresponding orthogonal dimension of the decoder output layer can participate in the softmax normalization calculation with its original value, and the projection value of the creative semantic attribute component in this orthogonal dimension can reach its maximum absolute value when unsuppressed, according to the natural output distribution of the creative strategy network. After restoring the complete generation capability, the creative strategy network can regenerate the full range of creative semantic attribute component values ​​in this orthogonal dimension.

[0094] Furthermore, before identifying creative semantic attribute components that exhibit coupling and resonance effects with audience response data across multiple cognitive dimensions based on system stability criterion values, the process also includes: Determine whether the data volume of the first advertising copy has reached the preset data volume threshold; When the amount of data delivered reaches the preset data volume threshold, the calculation of the system stability criterion value is triggered.

[0095] In this embodiment, the delivery data volume of the first advertising copy refers to the cumulative number of times the first advertising copy has been displayed from the time of delivery to the current time. The number of delivery impressions is obtained by the advertising platform by counting and accumulating the impressions each time the advertising copy is displayed to the audience.

[0096] In this embodiment, the preset data volume threshold refers to the minimum number of exposures required to determine whether the amount of data delivered for the first advertising copy is sufficient to reliably support the calculation of the system stability criterion. The preset data volume threshold is determined based on the parameter estimation accuracy requirements of the transfer functions in the transfer function matrix, and its value ranges from 500 to 2000. When the number of exposures is lower than the preset data volume threshold, the estimated values ​​of the gain coefficient, natural frequency, and damping ratio of each transfer function in the coupling relationship model have not yet converged to a statistically stable state. At this time, the calculated Nyquist proximity exhibits significant statistical fluctuations, and the determination result of the system stability criterion is unreliable.

[0097] In this embodiment, triggering the calculation of the system stability criterion value means initiating the process of drawing the Nyquist curve and calculating the Nyquist proximity of the closed-loop system when the amount of data delivered for the first ad copy reaches a preset data volume threshold. Before the amount of data delivered reaches the preset data volume threshold, the creative strategy network generates the first ad copy in its original state without applying suppression gain values ​​and continues to deliver it. At the same time, it continuously collects audience response data in four cognitive dimensions and updates the input-output sequence pairs until the amount of data delivered reaches the preset data volume threshold, at which point the first calculation of the system stability criterion value is performed.

[0098] This invention provides an embodiment of an advertising content generation system based on natural language processing, comprising: The Creative Strategy Network Module is used to construct a Creative Semantic Orthogonal Decoupling Space, which contains multiple orthogonal dimensions, each corresponding to an independent creative semantic attribute. The module maps the contextual signals of the advertising scenario to creative semantic attribute components in the Creative Semantic Orthogonal Decoupling Space, generating the first ad copy for delivery. The coupling relationship model construction module is used to obtain the audience's response data in multiple cognitive dimensions after the first advertising copy is delivered. Based on the relationship between the creative semantic attribute components and the audience's response data in multiple cognitive dimensions, a coupling relationship model between the creative semantic attribute components and the audience's response data in multiple cognitive dimensions is established. The resonance source identification module is used to identify creative semantic attribute components that have a coupling resonance effect with the audience's response data in multiple cognitive dimensions in the coupling relationship model based on the system stability criterion value, and to use them as resonance source semantic attribute components. The resonance suppression module is used to suppress the creative semantic attribute generation sub-path corresponding to the semantic attribute component of the resonance source in response to a new ad delivery request, and the creative strategy network module generates a second ad copy for delivery.

[0099] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of this invention and its equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for generating advertising content based on natural language processing, characterized in that, include: Construct a creative semantic orthogonal decoupling space, which contains multiple orthogonal dimensions, each of which corresponds to an independent creative semantic attribute; Deploy a creative strategy network to map the contextual signals of the advertising scenario into creative semantic attribute components in the creative semantic orthogonal decoupling space, and generate the first advertising copy for delivery; After the first ad copy is delivered, the audience's response data in multiple cognitive dimensions is obtained. Based on the relationship between the creative semantic attribute components and the audience's response data in multiple cognitive dimensions, a coupling relationship model between the creative semantic attribute components and the audience's response data in multiple cognitive dimensions is established. In the coupling relationship model, based on the system stability criterion, creative semantic attribute components that have a coupling resonance effect with the audience's response data in multiple cognitive dimensions are identified as the resonance source semantic attribute components. In response to a new ad delivery request, the creative semantic attribute generation sub-path corresponding to the semantic attribute component of the resonance source is suppressed, and the creative strategy network generates a second ad copy for delivery.

2. The advertising content generation method based on natural language processing according to claim 1, characterized in that, Constructing an orthogonal decoupling space for creative semantics, specifically including: Multiple orthogonal dimensions are defined, each corresponding to an independent creative semantic attribute. Creative semantic attributes include appeal type, emotional tone, language style, rhetorical devices, and narrative structure. Construct an orthogonal basis vector set, in which any two basis vectors are orthogonal to each other, and each basis vector corresponds to an orthogonal dimension; The creative semantic attribute components are generated by the creative policy network in the creative semantic orthogonal decoupled space spanned by the orthogonal basis vectors. The projection value of the creative semantic attribute component on each orthogonal dimension represents the activation intensity of the corresponding creative semantic attribute.

3. The advertising content generation method based on natural language processing according to claim 1, characterized in that, After obtaining the first ad copy, we analyzed the audience's response data across multiple cognitive dimensions, specifically including: The audience cognitive state space is constructed, which includes four cognitive dimensions: cognitive novelty seeking degree, cognitive fatigue accumulation degree, social identity sensitivity degree, and anchoring effect deviation degree. These four cognitive dimensions together constitute a four-dimensional representation of the audience's psychological schema characteristics. After the first ad copy is placed, collect user behavior data generated by the audience within a preset time window. User behavior data includes page dwell time, scroll depth, hotspot distribution of clicks, and number of times the ad was viewed. Feature extraction is performed on user behavior data to obtain user behavior feature vectors; the user behavior feature vectors are input into a preset behavior-cognition mapping model to calculate the response value corresponding to each cognitive dimension; the response values ​​corresponding to the four cognitive dimensions are combined to form the audience's response data on multiple cognitive dimensions.

4. The advertising content generation method based on natural language processing according to claim 3, characterized in that, Based on the relationship between creative semantic attribute components and audience response data across multiple cognitive dimensions, a coupling model is established between creative semantic attribute components and audience response data across multiple cognitive dimensions, specifically including: Obtain market feedback sequence data generated after the historical ad copy was launched. The historical ad copy refers to the ad copy that was launched before the first ad copy was launched. Combine the creative semantic attribute components of the historical ad copy with the audience's response data in four cognitive dimensions during the corresponding campaign period to form input-output sequence pairs. The input-output sequence pairs are systematically identified, and the gain coefficient, natural frequency, and damping ratio of the transfer function between each pair of creative semantic attribute components and the audience's response data in one of the four cognitive dimensions are estimated. The transfer function adopts the form of a second-order oscillatory element. The transfer function matrix is ​​composed of transfer functions. The element in the i-th row and j-th column of the transfer function matrix is ​​the transfer function between the i-th creative semantic attribute component and the j-th cognitive dimension. The transfer function matrix serves as a coupling relationship model between the creative semantic attribute component and the audience's response data in the four cognitive dimensions.

5. The advertising content generation method based on natural language processing according to claim 4, characterized in that, In the coupling relationship model, based on the system stability criterion, creative semantic attribute components that exhibit coupling resonance effects with the audience's response data across multiple cognitive dimensions are identified. These components specifically include: Using the creative strategy network as the forward pathway and the audience's response data in the four cognitive dimensions as the feedback pathway, a closed-loop system is constructed. The output of the creative strategy network is the creative semantic attribute component, and the audience's response data in the four cognitive dimensions is the feedback quantity of the closed-loop system. Plot the Nyquist curve of the closed-loop system on the complex plane, calculate the minimum value of the Euclidean distance from the complex coordinates corresponding to all sampling frequency points of the Nyquist curve to the preset critical point of the Nyquist criterion as the Nyquist proximity, and use the Nyquist proximity as the system stability criterion value. When the system stability criterion value is lower than the preset stability margin threshold, it is determined that there is a coupled resonance effect in the closed-loop system; In the transfer function matrix, transfer functions with a damping ratio less than a preset critical damping value and a gain coefficient exceeding a preset gain threshold are selected. The creative semantic attribute components corresponding to the selected transfer functions are identified as the semantic attribute components of the resonance source.

6. The advertising content generation method based on natural language processing according to claim 5, characterized in that, Suppression is applied to the creative semantic attribute generation sub-paths corresponding to the semantic attribute components of the resonance source, specifically including: Based on the system transfer function of the closed-loop system and the preset stability margin threshold, the inverse solution is the minimum gain attenuation required to restore the Nyquist proximity to above the preset stability margin threshold. The minimum gain attenuation is converted into the suppression gain value of the creative semantic attribute generation sub-path on the orthogonal dimension corresponding to the semantic attribute components of the resonance source; In the creative strategy network, a suppression gain value is applied to the creative semantic attribute generation sub-path on the orthogonal dimension corresponding to the semantic attribute component of the resonance source, thereby limiting the maximum intensity of activation of the corresponding orthogonal dimension by the creative strategy network in the subsequent generation process.

7. The advertising content generation method based on natural language processing according to claim 1, characterized in that, When acquiring audience response data across multiple cognitive dimensions after the first ad copy is placed, this also includes: Obtain the core advertising metrics set by the business partner. These core metrics include conversion rate and return on investment. The advertising conversion process is broken down into the exposure stage, click stage, page dwell stage, add-to-cart stage, and payment stage. Assign exposure completion rate to the exposure phase, click pass rate to the click phase, effective dwell time to the page dwell phase, add to cart conversion rate to the add to cart phase, and payment completion rate to the payment phase. Calculate the trend deviation between the core advertising metrics and each micro-conversion rate metric over time, and identify micro-conversion rate metrics whose trend deviation exceeds a preset deviation threshold as micro-conversion bottleneck metrics. Based on the micro-conversion bottleneck index, the audience cognitive decay index is extracted from the audience's response data across multiple cognitive dimensions. The audience cognitive decay index characterizes the rate at which the audience's response intensity to a specific creative semantic attribute decays with the number of repeated exposures.

8. The advertising content generation method based on natural language processing according to claim 6, characterized in that, Also includes: The creative semantic attribute generation sub-path with the applied suppression gain value is continuously monitored. The creative semantic attribute generation sub-path with the applied suppression gain value corresponds to the semantic attribute component of the resonance source. Continuously calculate the Nyquist proximity of the closed-loop system; When the Nyquist proximity recovers to above the preset recovery threshold, it is determined that the coupling resonance effect between the semantic attribute component of the resonance source and the audience's response data in the four cognitive dimensions has subsided. Remove the suppression gain value for the creative semantic attribute generation sub-path and restore the creative policy network's full generation capability in the corresponding orthogonal dimension.

9. The advertising content generation method based on natural language processing according to claim 1, characterized in that, Before identifying creative semantic attribute components that exhibit coupling and resonance effects with audience response data across multiple cognitive dimensions based on system stability criteria, the process also includes: Determine whether the data volume of the first advertising copy has reached the preset data volume threshold; When the amount of data delivered reaches the preset data volume threshold, the calculation of the system stability criterion value is triggered.

10. An advertising content generation system based on natural language processing, characterized in that, include: The Creative Strategy Network Module is used to construct a Creative Semantic Orthogonal Decoupling Space, which contains multiple orthogonal dimensions, each corresponding to an independent creative semantic attribute. The module maps the contextual signals of the advertising scenario to creative semantic attribute components in the Creative Semantic Orthogonal Decoupling Space, generating the first ad copy for delivery. The coupling relationship model construction module is used to obtain the audience's response data in multiple cognitive dimensions after the first advertising copy is delivered. Based on the relationship between the creative semantic attribute components and the audience's response data in multiple cognitive dimensions, a coupling relationship model between the creative semantic attribute components and the audience's response data in multiple cognitive dimensions is established. The resonance source identification module is used to identify creative semantic attribute components that have a coupling resonance effect with the audience's response data in multiple cognitive dimensions in the coupling relationship model based on the system stability criterion value, and to use them as resonance source semantic attribute components. The resonance suppression module is used to suppress the creative semantic attribute generation sub-path corresponding to the semantic attribute component of the resonance source in response to a new ad delivery request, and the creative strategy network module generates a second ad copy for delivery.

Images