An iterative optimization method and system based on generated content quality assessment feedback

Abstract

The application discloses an iterative optimization method and system based on generated content quality evaluation feedback, relates to the field of artificial intelligence, acquires advertisement creative task data, performs field recognition and constraint item extraction on the advertisement creative task data, classifies the constraint items, generates creative constraint information, generates quality evaluation configuration and initial candidate information stream advertisement scripts according to the creative constraint information, generates updated candidate information stream advertisement scripts according to the iterative optimization instruction, performs quality re-evaluation, candidate sorting and non-degeneration checking on the updated candidate information stream advertisement scripts, judges whether to continue iteration according to the non-degeneration checking result, judges the updated candidate information stream advertisement scripts in combination with the updated quality evaluation result and the quality performance of the last round of candidate scripts before output, so that the possibility that the quality of the scripts that have met the requirements is reduced after being changed in the iterative optimization process is reduced, and the stability of the target information stream advertisement script output process is improved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence, specifically to an iterative optimization method and system based on feedback from the quality assessment of generated content. Background Technology

[0002] In-feed advertising is a common form of online advertising. It typically displays text information such as ad titles, ad body, calls to action, and necessary prompts in in-feed ad slots. The effectiveness of in-feed ad copy is usually related to various factors such as product information, core selling points, target audience, advertising platform, advertising objectives, ad format, brand requirements, and restrictions. Different advertising platforms have different requirements for copy length, expression, display pace, and expression restrictions. Different advertising objectives also have different focuses on brand exposure, click-through rates, conversion rates, or lead generation. With the development of text generation models, the method of automatically generating in-feed ad copy based on ad creative task data is gradually being applied to ad creative production scenarios.

[0003] Existing ad copy generation methods typically generate candidate copy based on input creative task data and then evaluate these candidate copy as a whole using methods such as manual review, rule validation, or model scoring. In practical applications, the quality evaluation of in-feed ad copy usually involves multiple dimensions. For the same candidate copy, different copy units may perform differently under different quality assessment indicators. Simply scoring the candidate copy as a whole is insufficient to reflect the specific quality performance of different copy units. Furthermore, using the overall evaluation result as the evaluation standard may alter the expression content that already meets the requirements in the previous round of copy when iterating and rewriting the candidate copy, leading to a decrease in the quality of copy units that already meet brand expression, platform adaptation, or compliance and security requirements. Therefore, this paper proposes an iterative optimization method and system based on the quality assessment feedback of generated content to address the problems mentioned above. Summary of the Invention

[0004] To address the aforementioned technical issues, this paper provides an iterative optimization method and system based on the feedback of generated content quality assessment. This solution resolves the problem that existing advertising copy generation methods, as mentioned in the background, typically generate candidate copy based on input creative task data and then evaluate the candidate copy holistically using methods such as manual review, rule validation, or model scoring. However, in practical applications, the quality evaluation of feed ad copy usually involves multiple dimensions. For the same candidate copy, different copy units may perform differently under different quality assessment indicators. Simply scoring the candidate copy holist holistly is insufficient to reflect the specific quality performance of different copy units. Furthermore, using the overall evaluation result as the evaluation standard may alter the expression content that already meets the requirements in the previous round of copy during iterative rewriting of candidate copy, leading to a decrease in the quality of copy units that already meet brand expression, platform adaptation, or compliance and security requirements.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: An iterative optimization method based on generated content quality assessment feedback includes: Acquire advertising creative task data, which includes product information, core selling points, target audience information, advertising platform information, advertising target information, advertising layout information, brand information, and restriction information; Field identification and constraint extraction are performed on advertising creative task data, and the constraints are categorized to generate creative constraint information; Based on the creative constraint information, a quality assessment configuration and initial candidate information flow ad copy are generated. The quality assessment configuration includes a set of quality assessment indicators and the importance relationship of the indicators. The initial candidate in-feed ad copy was evaluated for copy unit quality, and the quality evaluation results were obtained. A dynamic quality benchmark is determined based on advertising creative task data, creative constraint information, and quality assessment results. A quality feedback vector is generated, an iterative optimization instruction is generated based on the quality feedback vector, and an updated candidate information flow ad copy is generated based on the iterative optimization instruction. The updated candidate information flow ad copy is re-evaluated for quality, ranked and non-degraded, and determined whether to continue iteration. The determination result is obtained, and the iteration instruction is regenerated or the updated candidate information flow ad copy is added to the output candidate set based on the determination result. Determine the target in-feed ad copy from the output candidate set, and output the target in-feed ad copy and the corresponding quality assessment results.

[0006] Furthermore, an iterative optimization system based on generated content quality assessment feedback is proposed to implement the iterative optimization method based on generated content quality assessment feedback as described above, including: The data management module is used to acquire advertising creative task data, identify fields and extract constraints from the advertising creative task data, classify the constraints, and generate creative constraint information. The generation control module is used to generate content generation control information based on creative constraint information, generate initial candidate information flow ad copy based on content generation control information, and generate updated candidate information flow ad copy based on iterative optimization instructions and the previous round of candidate information flow ad copy. The quality assessment module is used to generate quality assessment configuration based on the target information, platform information, and ad format information, and to perform copy unit quality assessment on the initial candidate information flow ad copy or the updated candidate information flow ad copy based on the quality assessment configuration, so as to obtain the quality assessment result or update the quality assessment result. The optimization decision module is used to determine a dynamic quality benchmark based on advertising creative task data, creative constraint information and quality assessment results, generate a quality feedback vector, and generate iterative optimization instructions based on the quality feedback vector; The iteration management module is used to sort candidates and perform non-degradation verification based on the update quality assessment results, and determine whether to continue iteration based on the non-degradation verification results. If iteration continues, the iteration instruction is regenerated; if iteration does not continue, the updated candidate information flow ad copy is added to the output candidate set. The output management module is used to determine the target information flow ad copy from the set of output candidates and output the target information flow ad copy and the corresponding quality assessment results.

[0007] The beneficial effects of this invention compared to the prior art are: This solution generates creative constraint information by identifying fields, extracting constraints, and classifying constraints from advertising creative task data. Based on this information, it generates quality assessment configurations and initial candidate in-feed ad copy. This allows the generation and evaluation of candidate ad copy to incorporate product information, core selling points, target audience information, advertising platform information, advertising objectives, ad format information, brand information, and limiting conditions, thereby improving the matching degree between candidate in-feed ad copy and advertising creative task data.

[0008] This solution assesses the quality of initial candidate in-feed ad copy by evaluating the quality of each copy unit. Based on the unit quality performance information of each copy unit, a quality score for the candidate copy is calculated. This ensures that the quality assessment results not only reflect the overall quality of the candidate copy but also the quality performance of different copy units under the corresponding quality assessment indicators, thereby improving the granularity of the quality evaluation results.

[0009] This solution determines a dynamic quality benchmark based on advertising creative task data, creative constraint information, and quality assessment results. It then generates a quality feedback vector based on the dynamic quality benchmark and generates iterative optimization instructions based on the quality feedback vector. This allows the iterative optimization process to combine the current advertising creative task and copy unit quality performance to determine the copy units to be retained, the copy units to be optimized, and the corresponding optimization actions, thereby improving the correspondence between copy optimization instructions and specific quality issues.

[0010] This solution generates updated candidate news feed ad copy based on iterative optimization instructions, and performs quality re-evaluation, candidate ranking, and non-degradation verification on the updated candidate news feed ad copy. Based on the non-degradation verification results, it determines whether to continue iterating. This allows the updated candidate news feed ad copy to be judged in conjunction with the updated quality evaluation results and the quality performance of the previous round of candidate copy before output. This reduces the possibility of quality degradation caused by changes to copy units that have already met the requirements during iterative optimization, and improves the stability of the target news feed ad copy output process. Attached Figure Description

[0011] Figure 1 This is a flowchart of an iterative optimization method based on generated content quality assessment feedback proposed in this invention; Figure 2 This is a flowchart illustrating the generation of creative constraint information in this invention; Figure 3 This is a flowchart of the quality assessment process in this invention; Figure 4 This is a flowchart illustrating the generation of updated candidate information stream ad copy in this invention; Figure 5 This is a structural block diagram of an iterative optimization system based on generated content quality assessment feedback proposed in this invention. Detailed Implementation

[0012] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.

[0013] Reference Figures 1-4 As shown, an iterative optimization method based on generated content quality assessment feedback includes: S1. Obtain advertising creative task data, which includes product information, core selling points, target audience information, advertising platform information, advertising target information, advertising layout information, brand information, and restriction information; Specifically, product information includes product name, product category, product function, and usage scenario. The product name is used to identify the specific product that the information flow advertising copy refers to. The product category is used to determine the industry or category to which the product belongs. The product function is used to indicate the purpose or function that the product can achieve. The usage scenario is used to indicate the specific environment, time, object, or demand scenario in which the product is used. The core selling points include the main selling points, auxiliary selling points, and selling point priority. The main selling points are used to express the core benefits that need to be emphasized in the information flow ad copy. Auxiliary selling points are used to supplement the main selling points. Selling point priority is used to determine the order, intensity, and content proportion of multiple selling points in the information flow ad copy. Target audience information includes audience profiles, consumption needs, pain points, interests and preferences, and purchase stage. Audience profiles represent the basic characteristics of the target audience, including age, gender, region, occupation, and spending power. Consumption needs represent the needs that the target audience wants to satisfy through the product. Pain points represent the problems that the target audience faces in the current usage or consumption scenario. Purchase stage represents the consumption stage that the target audience is in, such as awareness, interest, comparison, decision-making, or repurchase. The platform information includes the platform type and platform-restricted expressions. The platform type indicates the platform category on which the information flow ad copy is placed, and the platform-restricted expressions indicate the requirements of the corresponding platform for prohibited, restricted, or cautiously expressed content in the information flow ad copy.

[0014] The target information for ad placement includes exposure goals, click goals, conversion goals, lead generation goals, and brand awareness goals. Exposure goals are used to instruct the ad copy to focus on increasing impressions and brand visibility. Click goals are used to instruct the ad copy to focus on increasing users' willingness to click. Conversion goals are used to instruct the ad copy to focus on promoting purchases, inquiries, orders, or other conversion behaviors. Lead generation goals are used to instruct the ad copy to focus on guiding users to submit contact information or needs. Brand awareness goals are used to instruct the ad copy to focus on strengthening brand impressions, brand memorability, or brand image. The ad layout information indicates the text composition and display structure of the in-feed ad copy within the in-feed ad slot. This includes title layout information, body text layout information, title generation requirements, body text generation requirements, and content length requirements. Title layout information indicates the composition of the ad title. Body text layout information indicates the arrangement of the first sentence (attractive content), pain point expression content, product selling point content, benefit expression content, call to action content, and compliance reminder content within the ad body text. Title generation requirements limit the emphasis and format of the ad title. Body text generation requirements limit the content composition and order of the ad body text. Content length requirements limit the text length of the ad title and ad body text.

[0015] Brand information includes brand name, brand tone, brand keywords, and prohibited expressions. The brand name identifies the brand entity corresponding to the feed ad copy. Brand tone indicates the tone, style, and emotional inclination of the brand in the advertising expression. Brand tone includes required words and prohibited words. Required words indicate the tone words, style words, and habitual expressions that should be used in the brand copy. Prohibited words indicate the tone words, style words, and habitual expressions that are prohibited in the brand copy. Brand keywords indicate brand-related expressions that need to be reflected or prioritized in the feed ad copy. Prohibited expressions indicate brand-related statements that must not appear in the feed ad copy. The restrictions include word count limits, sensitive word limits, industry compliance limits, and style limits. Word count limits restrict the text length of the ad copy or its ad unit. Sensitive word limits restrict sensitive expressions that must not appear in the ad copy. Industry compliance limits restrict that the ad copy must comply with the compliance requirements of the relevant industry. Style limits restrict the boundaries of the expressive style of the ad copy.

[0016] S2. Perform field identification and constraint extraction on the advertising creative task data, classify the constraints, and generate creative constraint information; Furthermore, the advertising creative task data is processed for field identification and constraint extraction, and the constraints are categorized to generate creative constraint information, specifically including: S21. Extract product expression constraints and selling point expression constraints based on product information and core selling points; Specifically, product names, product categories, product functions, and usage scenarios are extracted from product information to form product expression constraints. These constraints are used to limit the target product, product scope, and usage scenarios that the information flow ad copy should point to, thus preventing candidate information flow ad copy from deviating from the actual target audience. Extract the main selling points, auxiliary selling points, and selling point priorities from the core selling points. These selling point expression constraints are then used to limit the key benefits, selling point order, and selling point expression scope that need to be emphasized in the information flow ad copy, so as to avoid the omission, weakening, or deviation of selling points in the candidate information flow ad copy.

[0017] S22. Extract audience expression constraints based on target audience information; Specifically, audience expression constraints are formed by extracting audience profiles, consumption needs, pain points, interests, and purchase stages from the target audience information. These constraints are used to limit the communication targets, pain point entry points, demand response directions, and expression methods of the information flow ad copy, so that the candidate information flow ad copy matches the information needs and acceptance habits of the target audience. S23. Extract platform format constraints based on the platform information; Specifically, platform type and platform-restricted expressions are extracted from the information of the advertising platform to form platform format constraints. Platform format constraints include requirements for the display format of the advertising platform, title length requirements, body text length requirements, requirements for avoiding platform-restricted expressions, and requirements for the display rhythm of the information flow. The requirements for the display rhythm of the information flow include the requirement that there should be transitional connecting words or referential relationships between adjacent copy units. Platform format constraints are used to limit the display format, text length, expression method, and prohibited content of information flow advertising copy in the advertising platform, so that the candidate information flow advertising copy adapts to the display rules and expression specifications of the advertising platform.

[0018] S24. Extract optimization direction constraints based on the delivery target information; Specifically, exposure goals, click goals, conversion goals, lead generation goals, and brand awareness goals are extracted from the target information. Optimization direction constraints are generated based on the enabled goals in the target information. These optimization direction constraints are used to determine the key directions for information flow ad copy during quality assessment and iterative optimization. For example, when the target audience is exposure, the optimization direction constraints focus on the brand visibility, information recognizability, and memorable expression of the in-feed ad copy. When the target audience is clicks, the optimization direction constraints focus on the attractiveness of the headline, the attractiveness of the first sentence, the ability to stimulate interest, and the expression of click guidance in the information flow ad copy. When the target audience is conversion, the optimization direction focuses on the clarity of the selling points, the expression of benefits, and the call to action in the copywriting of the in-feed ads. When the target audience is lead generation, the optimization focus is on conveying trust, matching needs, and providing easy-to-use guidance in the ad copy. When the target audience is brand awareness, the optimization focus is on brand consistency, brand keyword coverage, and expression of brand memorability in the in-feed ad copy.

[0019] S25. Extract copy structure constraints based on the advertisement layout information; Specifically, the copywriting structure constraints are formed by extracting title layout information, body text layout information, title generation requirements, body text generation requirements, and content length requirements from the advertisement layout information. These constraints include title structure constraints, first-sentence attractiveness structure constraints, pain point expression structure constraints, product selling point structure constraints, benefit point expression structure constraints, call to action structure constraints, and compliance reminder structure constraints. Title structure constraints limit the product target, main selling point, target audience, or attractive expressions in the advertisement title. First-sentence attractiveness structure constraints limit the interest-stimulating content or scenario introduction content in the first sentence of the advertisement body. Pain point expression structure constraints limit the approach to addressing the pain points of the target audience in the advertisement body. Product selling point structure constraints limit the expression of main and auxiliary selling points in the advertisement body. Benefit point expression structure constraints limit the expression of product benefits and use value in the advertisement body. Call to action structure constraints limit the expression guiding clicks, inquiries, purchases, or lead generation in the advertisement body. Compliance reminder structure constraints limit the risk warnings, restrictions, or compliance supplements in the advertisement body.

[0020] S26. Extract brand expression constraints based on brand information; Specifically, brand name, brand tone, brand keywords, and prohibited expressions are extracted from brand information to form brand expression constraints. These constraints include requirements for brand name usage, brand keyword coverage, brand tone, avoidance of prohibited expressions, and brand expression stability. Brand tone requirements include requirements for the use of brand tone keywords and avoidance of prohibited words. These constraints are used to limit the consistency between the in-feed ad copy and the brand image, brand tone, and brand expression boundaries, preventing candidate in-feed ad copy from deviating from the brand image or triggering brand expression risks.

[0021] S27. Extract risk boundary constraints based on the constraint information; Specifically, risk boundary constraints are formed by extracting word count limits, sensitive word limits, industry compliance limits, and style limits from the restrictive information. These risk boundary constraints include requirements for avoiding sensitive words, avoiding prohibited expressions by brands, avoiding prohibited expressions by platforms, avoiding prohibited expressions by industry compliance, avoiding exaggerated prohibited words, avoiding absolute prohibited words, and qualification prompts. Exaggerated prohibited words are used to identify exaggerated statements, and absolute prohibited words are used to identify absolute statements.

[0022] S28. Combine product expression constraints, selling point expression constraints, audience expression constraints, platform format constraints, optimization direction constraints, copywriting structure constraints, brand expression constraints, and risk boundary constraints to generate creative constraint information.

[0023] Specifically, the creative constraint information records the constraint content and constraint source field for each constraint item. The constraint source field is used to determine the source of quality problems during subsequent quality assessment, quality feedback, and iterative optimization.

[0024] S3. Generate quality assessment configurations and initial candidate in-feed ad copy based on creative constraint information; Specifically, the quality assessment configuration includes the set of quality assessment indicators and the importance relationships of the indicators; Furthermore, based on the creative constraint information, a quality assessment configuration and initial candidate in-feed ad copy are generated, specifically including: S31. Generate content generation control information based on creative constraint information. The content generation control information includes product target, main selling points, target audience, expression style, content structure, distribution platform, title generation requirements, body text generation requirements, content length requirements, and prohibited expression requirements. Specifically, the product name, product category, product function, and usage scenario are extracted from the product expression constraints to form the product object; the main selling points, auxiliary selling points, and selling point priorities are extracted from the selling point expression constraints to form the main selling points and their expression order; the target audience is formed by extracting the audience profile, consumption needs, pain points, and purchase stage from the audience expression constraints; the expression style is formed based on the brand tone, brand keywords, platform type, and platform restrictions; the content structure is formed based on the copywriting structure constraints; the placement platform is determined based on the platform format constraints; the title generation requirements, body text generation requirements, and content length requirements are formed based on the title generation requirements, body text generation requirements, and content length requirements in the ad layout information; and the prohibited expression requirements are formed based on the brand's prohibited expressions, sensitive word restrictions, industry compliance restrictions, and style restrictions.

[0025] S32. Generate multiple initial candidate in-feed ad copy based on content generation control information; Specifically, text generation task information is constructed based on content generation control information. This text generation task information includes product target, main selling points, auxiliary selling points, target audience, pain point information, expression style, content structure, advertising platform, title generation requirements, body text generation requirements, content length requirements, and prohibited expression requirements.

[0026] Furthermore, the text generation task information is input into the text generation model to generate multiple initial candidate information flow ad copy. The initial candidate information flow ad copy includes an ad title and an ad body. The ad title is used to express the product object, main selling points, target audience, or attractive content. The ad body is used to express the first sentence to attract attention, pain point expression, product selling points, benefit expression, call to action, and compliance reminders.

[0027] It should be noted that the text generation model is a pre-trained text generation model.

[0028] As a preferred implementation of this solution, the text generation model can be an autoregressive text generation model based on the GPT structure, or other pre-trained language models with natural language generation capabilities in the existing technology. The system constructs a model training sample set by acquiring information flow advertising copy samples, advertising creative task samples, and corresponding manually annotated copy quality results. Among them, the advertising creative task samples include product information, core selling points, target audience information, advertising platform information, advertising target information, advertising layout information, brand information, and restriction information. The information flow advertising copy samples include advertising title samples and advertising body samples. The manually annotated copy quality results include copy structure annotation, selling point expression annotation, audience matching annotation, brand expression annotation, platform adaptation annotation, and risk expression annotation. The system generates sample input text based on advertising creative task samples and outputs the corresponding feed ad copy samples. The sample input text is then fed into the initial text generation model, which generates predicted copy text word by word using an autoregressive generation method. The predicted copy text is compared with the sample output text to obtain the text generation error. Based on manually annotated copy quality results, constraints are imposed on the copy structure, selling point expression, audience matching, brand expression, platform adaptation, and risk expression in the predicted copy text to obtain the quality constraint error. The model parameters of the initial text generation model are updated based on the text generation error and the quality constraint error until both are less than a preset error value, such as 5%. The trained text generation model is then stored in a local model library and used to generate feed ad copy that meets the constraints of copy structure, brand expression, platform format, and risk boundary based on the text generation task information.

[0029] Furthermore, the text generation model generates an advertising title based on the product object, main selling points, target audience, expression style, title generation requirements, and content length requirements, and generates the advertising body text based on the target audience, pain point information, main selling points, auxiliary selling points, benefit expression requirements, call to action requirements, body text generation requirements, and prohibited expression requirements.

[0030] Furthermore, basic format checks are performed on the generated initial candidate information flow ad copy. These checks include title structure check, body structure check, content length check, brand prohibited expressions check, platform restricted expressions check, and sensitive word check. When the initial candidate information flow ad copy passes the basic format checks, it is recorded. When the initial candidate information flow ad copy fails the basic format checks, it is regenerated based on the failed checks.

[0031] Furthermore, for each initial candidate information flow ad copy, record the candidate number, generation round, and corresponding content generation control information. The generation round of the initial candidate information flow ad copy is recorded as round 1.

[0032] S33. Generate a set of quality assessment indicators and the importance relationships of the indicators based on the target information, platform information, and ad format information; Specifically, the quality assessment indicator set includes brand consistency indicators, audience matching indicators, selling point expression indicators, creative novelty indicators, conversion guidance indicators, platform adaptability indicators, content completeness indicators, and compliance and security indicators. When the campaign objective includes brand awareness goals, exposure goals, or when the ad creative task data includes brand name, brand tone, brand keywords, or prohibited brand expressions, activate the brand consistency indicator. When the campaign objective includes click goals, conversion goals, or lead generation goals, or when the ad creative task data includes audience profiles, consumer needs, pain points, or purchase stages, activate the audience matching indicator. When the ad creative task data includes main selling points, secondary selling points, or selling point priority, activate the selling point expression indicator. When the campaign objective includes exposure goals, click goals, or when the system generates multiple candidate feed ad copy or there is a previous round of candidate feed ad copy, activate the creative... The Novelty Index is activated when the campaign objective includes conversion or lead generation goals, or when the ad format includes calls to action, benefit reminders, or conversion guidance requirements. The Platform Adaptability Index is activated when the campaign platform information includes platform type, platform display format, platform format requirements, or platform-restricted expressions. The Content Completeness Index is activated when the ad format information includes requirements for title generation, body text generation, content length, or copywriting structure. The Compliance and Security Index is activated when the campaign platform information, brand information, or restriction information includes sensitive word restrictions, industry compliance restrictions, prohibited brand expressions, platform-restricted expressions, or other risks that cannot be output.

[0033] Among them, the brand consistency index is used to evaluate whether the candidate in-feed ad copy conforms to the brand expression requirements corresponding to the brand name, brand tone, brand keywords, and prohibited brand expressions; the audience matching index is used to evaluate whether the candidate in-feed ad copy matches the audience profile, consumption needs, pain points, and purchase stage in the target audience information; the selling point expression index is used to evaluate whether the candidate in-feed ad copy accurately, completely, and prominently expresses the main selling point and auxiliary selling point; and the creative novelty index is used to evaluate the degree of difference between the candidate in-feed ad copy and other candidate in-feed ad copy in the same round or the previous round. The conversion guidance index is used to evaluate whether candidate information flow ad copy has a clear call to action, benefit reminders, and conversion guidance. The platform compatibility index is used to evaluate whether candidate information flow ad copy conforms to the display format and expression restrictions of the platform. The content completeness index is used to evaluate whether candidate information flow ad copy covers the product target, core selling points, target audience, usage scenarios, and action guidance. The compliance and security index is used to evaluate whether candidate information flow ad copy contains sensitive words, brand prohibited expressions, platform restricted expressions, industry compliance restrictions, exaggerated prohibited words, absolute prohibited words, or other risks that cannot be output.

[0034] Furthermore, candidate quality assessment indicators that meet the activation conditions are added to the quality assessment indicator set to obtain the quality assessment indicator set corresponding to the current advertising creative task.

[0035] Furthermore, based on the hit rate of the activation conditions of the quality assessment indicators, the importance of each quality assessment indicator is calculated. For the j-th quality assessment indicator in the set of quality assessment indicators, its target hit value is determined. Hit value of the advertising platform and ad layout hit value , Specifically, when the activation condition of the j-th quality assessment indicator is matched by the target information, Take 1, otherwise Set to 0; When the activation condition of the j-th quality assessment indicator is matched by the information from the delivery platform, Take 1, otherwise Set to 0; When the activation condition of the j-th quality assessment indicator is matched by the advertisement layout information, Take 1, otherwise Set to 0; Furthermore, calculate the significance score of the j-th quality assessment indicator. : ; in, The base score, , , These are used to represent the score increase when the quality assessment indicator is matched by the target information, the platform information, and the ad format information, respectively. > > ; As a preferred embodiment of this solution, =1, =3, =2, =1. Since the target information directly reflects the core optimization direction of the current advertising creative task, the platform information reflects the platform display and publishing restrictions that the candidate information flow ad copy needs to meet, and the ad format information reflects the structure and length requirements of the candidate information flow ad copy, the target hit bonus, the platform hit bonus, and the ad format hit bonus are set to decrease in that order.

[0036] Furthermore, the sum of the importance scores of each quality assessment indicator in the quality assessment indicator set is calculated, and the proportion of the importance score of the j-th quality assessment indicator to the sum of the importance scores is determined as the importance of the j-th quality assessment indicator. : ; Where m represents the number of quality assessment indicators in the quality assessment indicator set. This indicates the relative importance of the j-th quality assessment indicator in the overall quality assessment.

[0037] Furthermore, the importance relationships of the indicators are formed by each quality assessment indicator and its corresponding degree of importance.

[0038] S4. Conduct copy unit quality assessment on the initial candidate in-feed ad copy and obtain the quality assessment results; Furthermore, the initial candidate in-feed ad copy is broken down into ad units. Based on the quality assessment configuration, each ad unit and the candidate in-feed ad copy as a whole are evaluated to generate quality assessment results, specifically including: S41. Based on the ad layout information and copy structure constraints, the initial candidate information flow ad copy is broken down into multiple copy units, which include headline unit, first sentence attraction unit, pain point expression unit, product selling point unit, benefit point expression unit, call to action unit, and compliance reminder unit. Specifically, based on the ad layout information and copy structure constraints, the initial candidate information flow ad copy is broken down into multiple copy units. If the initial candidate information flow ad copy does not contain independent compliance prompt text, the compliance prompt unit is set to an empty text unit.

[0039] Understandably, the system records the unit number, unit type, unit text, unit position, and the candidate in-feed ad copy for each ad unit.

[0040] S42. Based on the quality assessment index set, determine the unit quality performance information of each copywriting unit under the corresponding quality assessment index; Specifically, based on the quality assessment indicators in the quality assessment indicator set, the applicable evaluation items and scoring rules are determined for each copywriting unit. The unit quality performance information includes copywriting unit type, applicable quality assessment indicators, unit indicator score, unit quality score, unit hit content items, unit missing content items, and unit problem location. Furthermore, the evaluation items corresponding to the brand consistency index include brand name usage consistency evaluation items, brand keyword coverage evaluation items, brand tone matching evaluation items, brand prohibited expression avoidance evaluation items, and brand expression stability evaluation items; Specifically, when the candidate information feed ad copy contains a brand name and the brand name is exactly the same as the brand name in the brand information, the hit value of this evaluation item is 1; when the candidate information feed ad copy does not contain a brand name, or the brand name is different from the brand name in the brand information, the hit value of this evaluation item is 0. When the candidate news feed ad copy contains all the required brand keywords, the hit value of this evaluation item is 1; when the candidate news feed ad copy does not contain all the required brand keywords, the hit value of this evaluation item is 0. When the candidate news feed ad copy contains brand tone requirement words and there are no brand tone prohibition words, the hit value of this evaluation item is 1; when the candidate news feed ad copy does not contain brand tone requirement words or contains brand tone prohibition words, the hit value of this evaluation item is 0. When the candidate news feed ad copy does not contain any prohibited brand expressions, the hit value of this evaluation item is 1; when the candidate news feed ad copy contains any prohibited brand expressions, the hit value of this evaluation item is 0. When the brand name and brand keywords used in each copy unit of the candidate feed ad copy are consistent, the hit value of this evaluation item is 1; when the brand name or brand keywords used in different copy units of the candidate feed ad copy are inconsistent, the hit value of this evaluation item is 0.

[0041] Furthermore, the evaluation items corresponding to the audience matching index include target audience title matching evaluation items, consumption demand matching evaluation items, pain point information coverage evaluation items, purchase stage matching evaluation items, and usage scenario matching evaluation items. Specifically, when the candidate information feed ad copy contains the title specified in the target audience information, the hit value of this evaluation item is 1; when the candidate information feed ad copy does not contain the title specified in the target audience information, the hit value of this evaluation item is 0. When the candidate information feed ad copy contains demand keywords corresponding to consumer needs, the hit value of this evaluation item is 1; when the candidate information feed ad copy does not contain demand keywords corresponding to consumer needs, the hit value of this evaluation item is 0. When the candidate information feed ad copy contains the pain point keyword corresponding to the main pain point information, the hit value of this evaluation item is 1; when the candidate information feed ad copy does not contain the pain point keyword corresponding to the main pain point information, the hit value of this evaluation item is 0. When the candidate in-feed ad copy contains the guiding keyword corresponding to the current purchase stage, the hit value of this evaluation item is 1; when the candidate in-feed ad copy does not contain the guiding keyword corresponding to the current purchase stage, the hit value of this evaluation item is 0. When the candidate in-feed ad copy contains scenario keywords corresponding to the usage scenario, the hit value of this evaluation item is 1; when the candidate in-feed ad copy does not contain scenario keywords corresponding to the usage scenario, the hit value of this evaluation item is 0. Furthermore, the evaluation items corresponding to the selling point expression index include the main selling point coverage evaluation item, the main selling point highlighting evaluation item, the auxiliary selling point coverage evaluation item, the selling point priority consistency evaluation item, and the selling point and benefit point correspondence evaluation item. Specifically, when the candidate information flow ad copy contains all the main selling point keywords, the hit value of this evaluation item is 1; when the candidate information flow ad copy does not contain all the main selling point keywords, the hit value of this evaluation item is 0. When the main selling point keyword in the candidate information flow ad copy appears in the title unit or the first sentence attraction unit, the hit value of this evaluation item is 1; when the main selling point keyword in the candidate information flow ad copy does not appear in the title unit or the first sentence attraction unit, the hit value of this evaluation item is 0. When the candidate news feed ad copy contains all the required auxiliary selling point keywords, the hit value of this evaluation item is 1; when the candidate news feed ad copy does not contain all the required auxiliary selling point keywords, the hit value of this evaluation item is 0. When the order of the selling points in the candidate information flow ad copy is consistent with the order of priority of the selling points, the hit value of this evaluation item is 1; when the order of the selling points in the candidate information flow ad copy is inconsistent with the order of priority of the selling points, the hit value of this evaluation item is 0. When each key selling point keyword in the candidate news feed ad copy is followed by a benefit statement, the hit value of this evaluation item is 1; when there are key selling point keywords in the candidate news feed ad copy that are not followed by a benefit statement, the hit value of this evaluation item is 0.

[0042] Furthermore, the evaluation items corresponding to the creative novelty index include evaluation items for differences in title expression, differences in the first sentence expression, differences in copywriting structure, differences in the angle of selling point expression, and evaluation items for control of repetitive expression; Specifically, when the text repetition rate between the title of the candidate information flow ad copy and the comparison title is less than 50%, the hit value of this evaluation item is 1; when the text repetition rate between the title of the candidate information flow ad copy and the comparison title is not less than 50%, the hit value of this evaluation item is 0. When the text repetition rate between the first sentence of the candidate news feed ad copy and the first sentence of the comparison is less than 50%, the hit value of this evaluation item is 1; when the text repetition rate between the first sentence of the candidate news feed ad copy and the first sentence of the comparison is not less than 50%, the hit value of this evaluation item is 0. When the copy unit composition of the candidate information flow ad copy is different from that of the comparison copy and meets the copy structure constraints, the hit value of this evaluation item is 1; when the copy unit composition of the candidate information flow ad copy is the same as that of the comparison copy, the hit value of this evaluation item is 0. When the scenario keywords, benefit keywords, or pain point keywords used in the selling point expression of the candidate information flow ad copy are different from those in the comparison copy, the hit value of this evaluation item is 1; when the scenario keywords, benefit keywords, and pain point keywords used in the selling point expression of the candidate information flow ad copy are all the same as those in the comparison copy, the hit value of this evaluation item is 0. The hit value of this evaluation item is 1 when there are no three or more consecutive repeated phrases with the same words in the candidate information feed ad copy, and 0 when there are three or more consecutive repeated phrases with the same words in the candidate information feed ad copy.

[0043] Understandably, news feed ad titles are typically 10-30 characters long, and the first sentence is usually 15-40 characters long. A 50% duplication rate means that more than half of the characters are the same as the comparison text, which can be judged as a lack of expression. A duplication rate of less than 50% means that most of the content is a reorganized expression, which has expression differences. This value is taken as the middle dividing point, which can avoid normal keyword sharing being misjudged as duplication, and can also avoid being judged as having differences simply by replacing a few words. In Chinese expressions, two consecutive identical words may be considered normal rhetoric, such as "very, very" or "really, really". However, three or more consecutive identical words constitute obvious redundant and repetitive sentences. Therefore, using three as a dividing point can accurately distinguish between rhetorical repetition and redundant repetition.

[0044] Furthermore, the evaluation items corresponding to the conversion guidance index include evaluation items for clear benefit points, clear call to action, clear conversion path, expression of action threshold, and expression of trust enhancement. Specifically, when the candidate information flow ad copy contains benefit keywords corresponding to the main selling point, the hit value of this evaluation item is 1; when the candidate information flow ad copy does not contain benefit keywords corresponding to the main selling point, the hit value of this evaluation item is 0. When the candidate news feed ad copy contains a call to action, the hit value of this evaluation item is 1; when the candidate news feed ad copy does not contain a call to action, the hit value of this evaluation item is 0. When the candidate in-feed ad copy contains instructions on the user's next action path, the hit value of this evaluation item is 1; when the candidate in-feed ad copy does not contain instructions on the user's next action path, the hit value of this evaluation item is 0. When the candidate in-feed ad copy contains descriptions of low-threshold action conditions, the hit value of this evaluation item is 1; when the candidate in-feed ad copy does not contain descriptions of low-threshold action conditions, the hit value of this evaluation item is 0. When the candidate in-feed ad copy contains factual content such as data values, certification marks, user reviews, or the brand's establishment year, the hit value of this evaluation item is 1; when the candidate in-feed ad copy does not contain factual content such as data values, certification marks, user reviews, or the brand's establishment year, the hit value of this evaluation item is 0.

[0045] Furthermore, the evaluation items corresponding to the platform adaptability index include evaluation items for matching platform display format, title length compliance, body text length compliance, platform restriction expression avoidance, and information flow display rhythm adaptation. Specifically, when the composition of the candidate information flow ad copy is consistent with the display format required by the platform, the hit value of this evaluation item is 1; when the composition of the candidate information flow ad copy is inconsistent with the display format required by the platform, the hit value of this evaluation item is 0. When the number of characters in the title of the candidate news feed ad copy is within the required range for title length, the hit value of this evaluation item is 1; when the number of characters in the title of the candidate news feed ad copy exceeds the required range for title length, the hit value of this evaluation item is 0. When the number of characters in the body text of the candidate information flow ad copy is within the required range for body text length, the hit value of this evaluation item is 1; when the number of characters in the body text of the candidate information flow ad copy exceeds the required range for body text length, the hit value of this evaluation item is 0. When the candidate news feed ad copy does not contain any platform-restricted words, the hit value of this evaluation item is 1; when the candidate news feed ad copy contains any platform-restricted words, the hit value of this evaluation item is 0. When adjacent ad units in the candidate feed contain transitional connecting words or referential relationships, the hit value of this evaluation item is 1; when adjacent ad units in the candidate feed lack transitional connecting words and do not have referential relationships, the hit value of this evaluation item is 0.

[0046] Furthermore, the evaluation items corresponding to the content completeness indicator include product target coverage evaluation items, core selling point coverage evaluation items, target audience coverage evaluation items, usage scenario coverage evaluation items, pain point expression coverage evaluation items, benefit point expression coverage evaluation items, call to action coverage evaluation items, and compliance reminder coverage evaluation items. When the candidate in-feed ad copy contains product name or product category keywords, the hit value of this evaluation item is 1; when the candidate in-feed ad copy does not contain product name or product category keywords, the hit value of this evaluation item is 0. When the candidate information flow ad copy contains all the main selling point keywords and all the auxiliary selling point keywords, the hit value of this evaluation item is 1; when the candidate information flow ad copy does not contain all the main selling point keywords or does not contain all the auxiliary selling point keywords, the hit value of this evaluation item is 0. When the candidate news feed ad copy contains the target audience's title or characteristic words, the hit value of this evaluation item is 1; when the candidate news feed ad copy does not contain the target audience's title or characteristic words, the hit value of this evaluation item is 0. When the candidate in-feed ad copy contains scenario keywords corresponding to the usage scenario, the hit value of this evaluation item is 1; when the candidate in-feed ad copy does not contain scenario keywords corresponding to the usage scenario, the hit value of this evaluation item is 0. When the candidate information feed ad copy contains the pain point keyword corresponding to the main pain point information, the hit value of this evaluation item is 1; when the candidate information feed ad copy does not contain the pain point keyword corresponding to the main pain point information, the hit value of this evaluation item is 0. When the candidate information feed ad copy contains benefit keywords corresponding to the main selling point, the hit value of this evaluation item is 1; when the candidate information feed ad copy does not contain benefit keywords corresponding to the main selling point, the hit value of this evaluation item is 0. When the candidate news feed ad copy contains a call to action, the hit value of this evaluation item is 1; when the candidate news feed ad copy does not contain a call to action, the hit value of this evaluation item is 0. When a candidate news feed ad copy contains compliance keywords required by the industry, the hit value for this evaluation item is 1; when a candidate news feed ad copy does not contain compliance keywords required by the industry, the hit value for this evaluation item is 0.

[0047] Furthermore, the evaluation items corresponding to the compliance and security indicators include sensitive word hit evaluation items, brand prohibited expression hit evaluation items, platform restricted expression hit evaluation items, industry compliance restriction hit evaluation items, exaggerated prohibited word hit evaluation items, absolute prohibited word hit evaluation items, and qualification requirement fulfillment evaluation items. Specifically, when the candidate information flow ad copy does not contain any sensitive words in the sensitive word dictionary, the hit value of this evaluation item is 1; when the candidate information flow ad copy contains any sensitive words in the sensitive word dictionary, the hit value of this evaluation item is 0. When the candidate news feed ad copy does not contain any prohibited brand expressions, the hit value of this evaluation item is 1; when the candidate news feed ad copy contains any prohibited brand expressions, the hit value of this evaluation item is 0. When the candidate news feed ad copy does not contain any platform-restricted words, the hit value of this evaluation item is 1; when the candidate news feed ad copy contains any platform-restricted words, the hit value of this evaluation item is 0. When the candidate information feed ad copy does not contain any industry compliance restricted words, the hit value of this evaluation item is 1; when the candidate information feed ad copy contains any industry compliance restricted words, the hit value of this evaluation item is 0. When the candidate news feed ad copy does not contain any exaggerated prohibited words in the dictionary of prohibited words, the hit value of this evaluation item is 1; when the candidate news feed ad copy contains any exaggerated prohibited words in the dictionary of prohibited words, the hit value of this evaluation item is 0. When the candidate news feed ad copy does not contain any absolutely prohibited words in the absolute prohibited word dictionary, the hit value of this evaluation item is 1; when the candidate news feed ad copy contains any absolutely prohibited words in the absolute prohibited word dictionary, the hit value of this evaluation item is 0. When a candidate news feed ad copy contains product descriptions related to the qualification requirements of an industry and also includes corresponding qualification prompts, the hit value for this evaluation item is 1. When a candidate news feed ad copy contains product descriptions related to the qualification requirements of an industry but does not include corresponding qualification prompts, the hit value for this evaluation item is 0.

[0048] Furthermore, the internal importance of each evaluation item applicable to the same quality assessment indicator for the same copywriting unit is set to the same value. Let the u-th copywriting unit be applicable to n evaluation items under the j-th quality assessment indicator, then the internal importance of the k-th evaluation item is... for: ; Where k = 1, 2, ..., n, and the sum of the internal importance of each evaluation item of the u-th copywriting unit under the j-th quality assessment indicator is 1.

[0049] Let the applicable value of the u-th copywriting unit for the j-th quality assessment indicator be _____. When the j-th quality assessment indicator applies to the u-th copywriting unit, =1, when the j-th quality assessment indicator is not applicable to the u-th copywriting unit. =0, let the u-th copywriting unit include n evaluation items under the j-th quality assessment indicator, and the hit value of the k-th evaluation item is 0. The internal importance of the k-th evaluation item is: Then the unit indicator score of the u-th copywriting unit under the j-th quality assessment indicator. for: ; Where k = 1, 2, ..., n, Used to indicate whether the corresponding evaluation item was hit. Used to indicate the internal importance of the corresponding evaluation item; Furthermore, the system weights the unit indicator scores of the u-th copywriting unit under each quality assessment indicator according to the importance of each applicable quality assessment indicator, thus obtaining the unit quality score of the u-th copywriting unit. : ; Where j = 1, 2, ..., m, and m is the number of quality assessment indicators in the quality assessment indicator set. Let represent the importance of the j-th quality assessment indicator.

[0050] When the u-th copywriting unit has a missing evaluation item under the j-th quality assessment indicator, the system determines the copywriting unit position corresponding to the missing evaluation item as the unit problem position. When the u-th copywriting unit does not have a missing evaluation item under the j-th quality assessment indicator, the unit problem position is empty.

[0051] It should be noted that the hit rate of each evaluation item is determined by the semantic parsing model after performing semantic parsing on the candidate information flow ad copy, copy unit text, ad creative task data, and creative constraint information.

[0052] As a preferred implementation of this solution, the semantic parsing model can adopt a parallel combination model pre-built based on a rule dictionary, semantic matching model, text classification model, and risk identification model, or it can adopt other combination models with text semantic parsing capabilities in the existing technology. The system constructs a model training sample set by acquiring information flow advertising copy samples, advertising creative task samples, and corresponding manual annotation evaluation item hit results. Among them, the advertising creative task samples include product information, core selling points, target audience information, advertising platform information, advertising target information, advertising layout information, brand information, and restriction information. The information flow advertising copy samples include advertising title samples and advertising body samples. The manual annotation evaluation item hit results include brand consistency annotation, audience matching degree annotation, selling point expression degree annotation, creative novelty degree annotation, conversion guidance degree annotation, platform adaptability degree annotation, content completeness degree annotation, and compliance and security degree annotation, and each annotation result is represented by a hit value of 1 or 0. The rule dictionary is used to perform string matching on the text of copy units, and to identify brand name, brand keywords, prohibited brand expressions, brand tone requirement words, prohibited brand tone words, sensitive words, platform restricted expressions, industry compliance restricted expressions, prohibited exaggerated words, prohibited absolute words, target audience titles, pain point keywords, benefit point keywords, call to action guiding words, and compliance prompt keywords. The rule dictionary is pre-built by the system based on advertising creative task data, platform information, brand information, and restriction information and stored in the local rule base. The semantic matching model is used to calculate the semantic similarity between the copy unit text and the product object, core selling points, target audience, pain point information, usage scenarios and calls to action. When the semantic similarity is not lower than the preset similarity threshold, the corresponding evaluation item is determined to be matched. The semantic matching model can adopt a semantic encoding model based on a dual-tower structure, which encodes the copy unit text and the reference text into semantic vectors and then calculates the cosine similarity. The text classification model is used to identify the expressive intent and content category of copy units, classifying them into one of the following: headline expression, first sentence attraction, pain point expression, product selling point, benefit expression, call to action, or compliance reminder. The text classification model can use a text classifier based on a pre-trained language model and be fine-tuned by using samples labeled with the copy unit category. The risk identification model is used to identify whether there are exaggerated prohibited words, absolute prohibited words, expressions of lack of qualifications, and other risks that cannot be output in the text of the copy unit. The risk identification model can adopt a sequence labeling model based on a pre-trained language model and be fine-tuned by using samples labeled with the location of risk expressions and risk categories. The four sub-models work in parallel and collaboratively. The rule dictionary outputs the hit results of each prohibited word and keyword, the semantic matching model outputs the semantic similarity calculation results, the text classification model outputs the copywriting unit category determination results, and the risk identification model outputs the risk expression identification results. After summarizing the above output results, the system determines the hit value of each evaluation item according to the hit determination conditions of each evaluation item. When the hit determination conditions of an evaluation item are met, the hit value of that evaluation item is determined to be 1. When the hit determination conditions of an evaluation item are not met, the hit value of that evaluation item is determined to be 0. Finally, the semantic matching results between the copywriting unit and each evaluation item are obtained.

[0053] S43. Calculate the candidate copy quality score based on the unit quality performance information of each copy unit; Specifically, the candidate copy quality score of the candidate in-feed ad copy is calculated based on the unit quality score of each copy unit. Let the initial candidate in-feed ad copy include r copy units, and the unit quality score of the u-th copy unit be... The candidate copywriting quality score C is: ; Where r is the number of copy units actually included in the initial candidate feed ad copy.

[0054] S44. Generate quality assessment results based on unit quality performance information and candidate copy quality scores.

[0055] Specifically, the quality assessment results include the unit quality performance information and candidate copy quality scores for each copy unit. The unit quality performance information for each copy unit includes the copy unit type, applicable quality assessment indicators, unit indicator scores, unit quality scores, hit content items, missing content items, and unit problem locations.

[0056] S5. Determine the dynamic quality benchmark based on the advertising creative task data, creative constraint information and quality assessment results, generate a quality feedback vector, generate iterative optimization instructions based on the quality feedback vector, and generate updated candidate information flow advertising copy based on the iterative optimization instructions; Furthermore, a dynamic quality benchmark is determined based on advertising creative task data, creative constraint information, and quality assessment results, and a quality feedback vector is generated. Iterative optimization instructions are then generated based on the quality feedback vector, and updated candidate in-feed ad copy is generated based on these instructions. Specifically, this includes: S51. Determine dynamic quality benchmarks based on quality assessment results; Specifically, a dynamic quality benchmark is determined based on the unit indicator scores of each copywriting unit, thus establishing the dynamic quality benchmark for the u-th copywriting unit under the j-th quality assessment indicator. ; When the current candidate news feed ad copy is a candidate news feed ad copy from the first round, the system determines the average of the unit indicator scores of all ad units in the same generation round that have the same ad unit type as the u-th ad unit and are subject to the j-th quality evaluation indicator as the dynamic quality benchmark of the u-th ad unit under the j-th quality evaluation indicator. ; Furthermore, when the current candidate news feed ad copy is not a candidate news feed ad copy from the first round, obtain the unit indicator scores of each ad copy unit in the same generation round that has the same ad copy unit type as the u-th ad copy unit and is subject to the j-th quality evaluation indicator, and calculate their average value to obtain the unit average of the current round. Suppose that in the same generation round, there are N copywriting units with the same copywriting unit type as the u-th copywriting unit and applicable to the j-th quality assessment indicator, and the unit indicator score of the q-th copywriting unit under the j-th quality assessment indicator is: ; ; Where q = 1, 2, ..., N.

[0057] Obtain the quality assessment results corresponding to the candidate feed ad copy from the previous round, and read the unit indicator score of the historical copy unit corresponding to the u-th copy unit under the j-th quality assessment indicator from the previous round's quality assessment results, and use it as the historical unit score. ; Among them, the historical copy unit corresponding to the u-th copy unit is determined according to the copy unit type, copy unit number and copy structure constraints. When there is a historical copy unit in the previous quality assessment results that is consistent with the copy unit type and copy unit number of the u-th copy unit, the historical copy unit is determined as the corresponding historical copy unit. When there is no historical copy unit that is consistent with the copy unit type and copy unit number, but there is a historical copy unit with the same copy unit type, the historical copy unit with the same copy unit type is determined as the corresponding historical copy unit. When historical unit scores exist At that time, the system calculates based on the current round unit average. and historical unit rating Determine dynamic quality benchmark : ; When there is no historical unit score At that time, the average value of the current round units will be... Determined as a dynamic quality benchmark .

[0058] S52. Determine the relative quality differences of units based on the quality assessment results and dynamic quality benchmarks; Specifically, based on the unit indicator score of the u-th copywriting unit under the j-th quality assessment indicator. and dynamic quality benchmark Calculate the relative quality difference of the u-th copywriting unit under the j-th quality assessment index. : ; Among them, when When the value is greater than 0, it indicates that the u-th copywriting unit is below the dynamic quality benchmark under the j-th quality assessment indicator. When the value equals 0, it indicates that the u-th copywriting unit has reached the dynamic quality benchmark under the j-th quality assessment indicator. When the value is less than 0, it means that the u-th copywriting unit is higher than the dynamic quality benchmark under the j-th quality assessment indicator.

[0059] S53. Determine optimization priorities based on the relative quality differences of the units; Specifically, the system is based on the relative mass difference of the units. The importance of the j-th quality assessment indicator Determine the optimization priority of the u-th copywriting unit under the j-th quality assessment indicator. : ; in, Obtained from the importance relationships of indicators in the quality assessment configuration. This is used to indicate the optimization priority of the u-th copywriting unit under the j-th quality assessment indicator. The larger the value, the more priority the copywriting unit needs to be optimized under this quality assessment indicator.

[0060] S54. Generate a quality feedback vector based on the unit quality performance information and optimization priority; Specifically, the quality feedback vector includes copy unit type, applicable quality assessment indicators, unit indicator score, unit quality score, dynamic quality benchmark, relative quality difference of units, optimization priority, hit content items of units, missing content items of units, unit advantage tags, unit defect tags, and unit problem location; Among them, the copywriting unit type, applicable quality assessment indicators, unit indicator score, unit quality score, unit hit content items, unit missing content items, and unit problem location are derived from the unit quality performance information in the quality assessment results, and the dynamic quality benchmark is determined by S51; the relative quality difference of units is determined by S52, and the optimization priority is determined by S53.

[0061] Furthermore, when the u-th copywriting unit scores the unit indicator score under the j-th quality assessment indicator... Greater than or equal to the dynamic quality benchmark When the system marks the u-th copywriting unit as a unit advantage under the j-th quality assessment indicator, the unit indicator score of the u-th copywriting unit under the j-th quality assessment indicator is... Less than the dynamic quality benchmark At that time, the system will mark the u-th copywriting unit as a unit defect under the j-th quality assessment indicator.

[0062] S55. Determine the copy units to be retained and the copy units to be optimized based on the quality feedback vector; Specifically, when the u-th copy unit has no unit defect item, no missing content item, and no unit problem location, the system will determine the u-th copy unit as a copy unit to be retained; When the u-th copywriting unit has a defective item, a missing content item, or a problematic location, the system will identify the u-th copywriting unit as a copywriting unit to be optimized.

[0063] S56. Generate optimization actions based on the text unit to be optimized; Specifically, when a copy unit to be optimized has brand consistency defects, the system generates brand expression adjustment actions; Understandably, this is used to instruct the text generation model to supplement or correct brand names in copywriting units, add missing brand keywords, replace prohibited brand tone words with required brand tone words, delete prohibited brand expressions, and unify the use of brand names and brand keywords in all copywriting units.

[0064] When the copy unit to be optimized has audience matching defects, the system generates audience expression matching actions; Understandably, this is used to instruct the text generation model to supplement the copy unit with terms of address for the target audience, demand keywords corresponding to consumption needs, pain point keywords corresponding to the main pain point information, guiding words corresponding to the current purchase stage, and scenario keywords corresponding to the usage scenario.

[0065] When a copy unit to be optimized has a deficiency in the expression of selling points, the system generates actions to strengthen those selling points. Understandably, this is used to instruct the text generation model to supplement missing main selling point keywords and auxiliary selling point keywords in the copywriting unit, adjust the main selling point keywords to the title unit or the first sentence attraction unit, adjust the order of selling point expression to be consistent with the selling point priority, and supplement the corresponding benefit point expression content for main selling points that are not followed by benefit points.

[0066] When a copywriting unit to be optimized has a lack of creative originality, the system generates a differentiated expression action. Understandably, this is used to instruct the text generation model to adjust the wording and sentence structure of the title and first sentence to reduce the repetition rate with the comparison text, adjust the composition of the copy units to form a different structure from the comparison copy, change the scene keywords, benefit keywords or pain point keywords used in the expression of selling points, and delete repeated sentences composed of consecutive identical words.

[0067] When a copy unit to be optimized has a defect in conversion guidance, the system generates conversion guidance enhancement actions; Understandably, this is used to instruct the text generation model to supplement the copy unit with keywords of benefits corresponding to the main selling point, call to action prompts, descriptions of the user's next action path, descriptions of low-threshold action conditions, as well as factual content such as data values, certification marks, user reviews, or the brand's establishment years.

[0068] When a copy unit to be optimized has platform compatibility defects, the system generates a platform compatibility adjustment action. Understandably, this is used to instruct the text generation model to adjust the composition of copy units to match the display format of the delivery platform, delete or supplement text so that the number of characters in the title and body text falls within the corresponding length requirement range, delete platform-restricted expressions, and add transitional connecting words or establish referential relationships between adjacent copy units.

[0069] When a copy unit to be optimized has content incompleteness defects, the system generates a content completion action; Understandably, this is used to instruct the text generation model to supplement the copy unit with product name or product category keywords, missing main selling point keywords or auxiliary selling point keywords, target audience appellations or target audience characteristic words, scenario keywords corresponding to the usage scenario, pain point keywords corresponding to the main pain point information, benefit point keywords corresponding to the main selling point, call to action guiding words, and compliance prompt keywords required by the industry.

[0070] When a copywriting unit to be optimized has compliance and security deficiencies, the system generates a compliance correction action.

[0071] Understandably, this is used to instruct the text generation model to delete sensitive words, brand-prohibited expressions, platform-restricted expressions, industry-compliant restricted expressions, exaggerated prohibited words, and absolute prohibited words in the copywriting unit, or replace them with compliant synonyms, and to supplement the product expressions involving industries with qualification requirements with corresponding qualification prompts.

[0072] Furthermore, when a unit of copywriting to be optimized has missing content items, the system generates content supplementation actions based on the missing content items. Understandably, this is used to instruct the text generation model to supplement the missing content category in the appropriate position of the copy unit based on the missing content category recorded in the unit's missing content item. The supplementary content comes from the corresponding category data in the advertising creative task data and creative constraint information.

[0073] When a unit of copy to be optimized has a problem location, the system generates a partial rewriting action based on the problem location.

[0074] It is understandable that this is used to instruct the text generation model to modify only the text fragment corresponding to the unit problem location, while the text content outside the unit problem location remains unchanged, and the rewriting direction is determined according to the type of defect that caused the problem location.

[0075] S57. Generate iterative optimization instructions based on the text units to be retained, the text units to be optimized, and the optimization actions; Specifically, the iterative optimization instructions include the copy units to be retained, the copy units to be optimized, the quality assessment indicators to be optimized, the optimization actions, the location of the unit issues, and the content that should not be changed. Among them, the copy units to be retained and the copy units to be optimized are determined by S55, the quality assessment indicators to be optimized are determined according to the quality assessment indicators corresponding to the unit defect items, the optimization actions are determined by S56, the unit problem location comes from the unit quality performance information, and the prohibited changes to the content are determined according to the copy units to be retained, the content generation control information, and the creative constraint information.

[0076] It should be noted that when the same copywriting unit to be optimized corresponds to multiple optimization actions, the system will sort the multiple optimization actions according to the priority order of compliance correction actions, platform adaptation adjustment actions, brand expression adjustment actions, content completion actions, selling point enhancement actions, audience expression matching actions, conversion guidance enhancement actions, and expression differentiation actions. When there are conflicts between multiple optimization actions, the constraints corresponding to compliance correction actions, platform adaptation adjustment actions, and prohibited content changes will be satisfied first, and other optimization actions will be executed within the remaining modifiable scope.

[0077] S58. Generate updated candidate news feed ad copy based on iterative optimization instructions; Specifically, the system will iteratively optimize the instructions, content generation control information, and the input text generation model of the previous round of candidate information flow ad copy to generate updated candidate information flow ad copy; Among them, the content generation control information is generated by S31 based on the creative constraint information, and the candidate information flow ad copy of the previous round is the candidate information flow ad copy that has been generated and completed the quality assessment before the current iteration round.

[0078] Furthermore, the text generation model retains the main content of the text unit to be retained according to the iterative optimization instructions, and rewrites, supplements or adjusts the content of the text unit to be optimized according to the optimization actions to obtain updated candidate information flow ad copy.

[0079] S6. Perform quality review, candidate sorting and non-degradation verification on the updated candidate information flow ad copy, determine whether to continue iteration, obtain the judgment result, and regenerate the iteration instruction or add the updated candidate information flow ad copy to the output candidate set based on the judgment result. Furthermore, the updated candidate news feed ad copy undergoes quality re-evaluation, candidate ranking, and non-degradation verification to determine whether to continue iteration. Based on the judgment result, iteration instructions are regenerated or the updated candidate news feed ad copy is added to the output candidate set. Specifically, this includes: S61. Conduct a quality review of the updated candidate information flow ad copy to obtain the updated quality assessment results; Specifically, S4 is re-executed on the updated candidate feed ad copy according to the quality assessment configuration to obtain the updated quality assessment results. The updated quality assessment results include the unit quality performance information and the updated candidate ad copy quality score for each copy unit in the updated candidate feed ad copy.

[0080] S62. Calculate candidate ranking values ​​based on the updated quality assessment results; Specifically, the candidate ranking value for updating candidate in-feed ad copy is calculated based on the quality score, quality improvement value, and retention value of the units to be retained. Let the candidate ranking value of the i-th updated candidate news feed ad copy be... Update the candidate copywriting quality score to Quality improvement value The value to be retained for the cell is Then the candidate sorting value for: ; Among them, the quality score of the updated candidate copywriting was updated. Quality improvement value from updated quality assessment results To update the difference between the candidate copy quality score and the previous round's candidate copy quality score, a value of 0 is used when the difference is less than 0. The previous round's candidate copy quality score comes from the quality assessment results corresponding to the previous round's candidate feed ad copy. The value to be retained for the unit is set as follows. The retainable copy units are determined by S55 based on their retention status in the updated candidate feed ad copy. Suppose that the iterative optimization instruction includes L text units to be retained, among which... If each unit to be retained retains its original unit type and main content in the updated candidate feed ad copy, then the value of the unit to be retained will be maintained. for: ; When there are no text units to be retained in the iterative optimization instructions Take 1.

[0081] S63. Generate candidate sorting results based on candidate sorting values; Specifically, each candidate news feed ad copy is sorted in descending order based on its candidate ranking value to obtain the candidate ranking result. At least one candidate news feed ad copy ranked first in the candidate ranking result is identified as the candidate news feed ad copy to be verified.

[0082] S64. Perform non-degradation validation on the candidate information flow ad copy to be validated, and obtain the non-degradation validation result; Specifically, based on the quality assessment results of the previous round of candidate feed ad copy and the updated quality assessment results of the candidate feed ad copy to be verified, non-degradation verification is performed on the candidate feed ad copy to be verified.

[0083] The system determines that a candidate in-feed ad copy passes non-degenerate validation if it meets the following conditions: The candidate copy quality score of the candidate in-feed ad copy to be verified shall not be lower than the candidate copy quality score of the previous round. The text units to be retained have not been lost; The main content of the text units to be retained has not changed; The unit quality score of the copywriting unit to be retained is not lower than the unit quality score of the corresponding copywriting unit in the previous round.

[0084] Among them, the candidate copy quality score of the candidate information flow ad copy to be verified comes from the updated quality assessment results, the previous round candidate copy quality score comes from the quality assessment results corresponding to the previous round candidate information flow ad copy, the copy unit to be retained is determined by S55, and the unit quality score of the copy unit to be retained comes from the unit quality performance information of the corresponding round.

[0085] When the candidate copy quality score of the candidate in-feed ad copy to be verified is lower than the candidate copy quality score of the previous round, or the copy unit to be retained is lost, or the main content of the copy unit to be retained has changed, or the unit quality score of the copy unit to be retained is lower than the unit quality score of the corresponding copy unit in the previous round, the system determines that it has failed the non-degenerate verification.

[0086] S65. Determine whether to continue iteration based on the non-degenerate verification result. If it continues iteration, regenerate the iteration instruction. If it does not continue iteration, add the updated candidate information flow ad copy to the output candidate set. Specifically, when a top-ranked candidate news feed ad copy passes the non-degenerate validation and meets the iteration stop condition, the system stops iterating and adds the candidate news feed ad copy to the output candidate set.

[0087] When a top-ranked candidate news feed ad copy passes the non-degenerate validation but does not meet the iteration stop condition, the S5 process is re-executed based on the current round of update quality assessment results to generate a new iteration optimization instruction. When the top-ranked candidate news feed ad copy fails the non-degradation check and does not meet the iteration stop condition, the system determines the reason for quality degradation based on the non-degradation check result and regenerates the iteration optimization instruction based on the reason for quality degradation.

[0088] The reasons for quality degradation include one or more of the following: a decrease in the quality score of candidate copywriting, loss of copywriting units to be retained, changes in the expression of copywriting units to be retained, and a decrease in the quality score of copywriting units to be retained. When a candidate news feed ad copy ranked high fails the non-degenerate validation and meets the iteration stop condition, the candidate news feed ad copy with the highest candidate ranking value is selected from the candidate news feed ad copy that has passed the non-degenerate validation in the historical iteration process and added to the output candidate set. When there is no candidate news feed ad copy that has passed the non-degenerate validation in the historical iteration process, the system generates a prompt message that no output candidate news feed ad copy has been obtained.

[0089] The iteration stopping conditions include one or more of the following: the candidate copy quality score reaches the preset target score, the current iteration number reaches the preset maximum iteration number, and the quality improvement value is less than the preset improvement threshold.

[0090] It should be noted that the preset target score is the minimum quality score that candidate feed ad copy should achieve when it meets the output requirements. It is preset by the system. The preset maximum number of iterations is the maximum number of optimization rounds for the same candidate feed ad copy. It is preset by the system. The preset improvement threshold indicates whether there is still effective room for quality improvement when the quality improvement value of two adjacent rounds of candidate feed ad copy is less than the preset improvement threshold. It means that the quality improvement brought by continuing iteration is small, and the system can stop continuing iteration. The preset target score, preset maximum number of iterations, and preset improvement threshold can all be configured according to the quality requirements of the generated copy.

[0091] S7. Determine the target information flow ad copy from the output candidate set, and output the target information flow ad copy and the corresponding quality assessment results.

[0092] Furthermore, the target in-feed ad copy is determined from the output candidate set, and the target in-feed ad copy and corresponding quality assessment results are output, specifically including: S71. Determine the target in-feed ad copy from the set of output candidates; Specifically, from the output candidate set that has passed the non-degenerate validation, the updated candidate news feed ad copy with the highest candidate ranking value is selected as the target news feed ad copy. When there is only one updated candidate news feed ad copy in the output candidate set, the system determines that updated candidate news feed ad copy as the target news feed ad copy.

[0093] S72. Output the target information flow ad copy; Specifically, the system outputs target information flow ad copy. The target information flow ad copy includes the target ad title and the target ad body. The target ad body includes one or more of the following: an attractive opening sentence, a pain point statement, product selling points, benefits, a call to action, and compliance tips.

[0094] S73. Output the quality assessment results corresponding to the target information flow ad copy; Specifically, the output includes the quality assessment results corresponding to the target information flow ad copy. These results include the candidate ad copy quality scores, unit quality performance information for each ad copy unit, candidate ranking results, and iteration records. The unit quality performance information for each ad copy unit includes the ad copy unit type, applicable quality assessment indicators, unit indicator scores, unit quality scores, hit content items, missing content items, unit strength tags, unit weakness tags, and unit problem locations. The iteration records include candidate information flow ad copy for each round, quality assessment results for each round, dynamic quality benchmarks for each round, quality feedback vectors for each round, iteration optimization instructions for each round, candidate ranking results for each round, and non-degradation verification results for each round.

[0095] Reference Figure 5 As shown, an iterative optimization system based on generated content quality assessment feedback is proposed to implement the iterative optimization method based on generated content quality assessment feedback as described in any of the preceding claims, including: The data management module is used to acquire advertising creative task data, identify fields and extract constraints from the advertising creative task data, classify the constraints, and generate creative constraint information. Specifically, the data management module is used to structure the data of advertising creative tasks, identify the content corresponding to each field, and extract the constraints related to the generation, quality assessment and iterative optimization of information flow advertising copy from each field, so that the subsequent copy generation process can be restricted by creative constraint information.

[0096] The generation control module is used to generate content generation control information based on creative constraint information, generate initial candidate information flow ad copy based on content generation control information, and generate updated candidate information flow ad copy based on iterative optimization instructions and the previous round of candidate information flow ad copy. Specifically, when generating initial candidate information flow ad copy, the generation control module determines the content of the ad title and ad body based on the content generation control information. When generating updated candidate information flow ad copy, it retains the copy units to be retained according to the iterative optimization instructions, and performs partial rewriting, content supplementation or expression adjustment on the copy units to be optimized.

[0097] The quality assessment module is used to generate quality assessment configuration based on the target information, platform information, and ad format information, and to perform copy unit quality assessment on the initial candidate information flow ad copy or the updated candidate information flow ad copy based on the quality assessment configuration, so as to obtain the quality assessment result or update the quality assessment result. Specifically, the quality assessment module breaks down the candidate news feed ad copy into ad units based on the quality assessment configuration, and determines the unit quality performance information of each ad unit under the corresponding quality assessment indicators. The quality assessment results are updated by updating the unit quality performance information of each ad unit in the candidate news feed ad copy and updating the candidate ad copy quality score.

[0098] The optimization decision module is used to determine a dynamic quality benchmark based on advertising creative task data, creative constraint information and quality assessment results, generate a quality feedback vector, and generate iterative optimization instructions based on the quality feedback vector; Specifically, the optimization decision module is used to determine the relative quality difference of each copywriting unit based on the quality assessment results and dynamic quality benchmarks, and to determine the optimization priority based on the relative quality difference of the units. At the same time, it determines the content to be retained and the content to be optimized based on the quality feedback vector, and generates corresponding iterative optimization instructions.

[0099] The iteration management module is used to sort candidates and perform non-degradation verification based on the update quality assessment results, and determine whether to continue iteration based on the non-degradation verification results. If iteration continues, the iteration instruction is regenerated; if iteration does not continue, the updated candidate information flow ad copy is added to the output candidate set. Specifically, the iteration management module is used to calculate the candidate ranking value of the updated candidate information flow ad copy based on the update quality assessment results, and generate the candidate ranking result based on the candidate ranking value. The iteration management module is also used to perform non-degenerate verification on the candidate information flow ad copy to be verified based on the quality assessment results corresponding to the candidate information flow ad copy in the previous round and the update quality assessment results corresponding to the candidate information flow ad copy to be verified, so as to determine whether the updated candidate information flow ad copy has a decrease in the quality score of the candidate copy, the copy unit to be retained is lost, the expression of the copy unit to be retained has changed, or the quality score of the copy unit to be retained has decreased. At the same time, based on the non-degenerate verification results and the iteration stopping condition, it determines whether to continue iterating, and determines whether to add the updated candidate information flow ad copy to the output candidate set based on the judgment result.

[0100] The output management module is used to determine the target information flow ad copy from the set of output candidates and output the target information flow ad copy and the corresponding quality assessment results.

[0101] Specifically, the output management module selects updated candidate in-feed ad copy that meets the output requirements from the set of output candidates as the target in-feed ad copy, and outputs the quality evaluation results corresponding to the target in-feed ad copy.

[0102] Furthermore, the data management module includes: The data recognition unit is used to acquire and identify product information, core selling points, target audience information, advertising platform information, advertising target information, advertising layout information, brand information, and restriction information in advertising creative task data; The constraint processing unit is used to extract product expression constraints, selling point expression constraints, audience expression constraints, platform format constraints, optimization direction constraints, copywriting structure constraints, brand expression constraints, and risk boundary constraints from advertising creative task data, and to classify each constraint to generate creative constraint information; Furthermore, the generation control module includes: A configuration unit is generated to generate content generation control information based on creative constraint information. The candidate generation unit is used to generate initial candidate information flow ad copy based on content generation control information, and to generate updated candidate information flow ad copy based on iterative optimization instructions, content generation control information, and the previous round of candidate information flow ad copy.

[0103] Furthermore, the quality assessment module includes: The evaluation configuration unit is used to generate a set of quality evaluation indicators and the importance relationships of the indicators based on the target information, the platform information, and the ad format information. The unit evaluation unit is used to break down candidate infeed ad copy into multiple copy units, determine the unit quality performance information of each copy unit based on the quality assessment indicator set and the importance relationship of the indicators, calculate the candidate copy quality score, and generate quality assessment results. Furthermore, the optimization decision-making module includes: The quality analysis unit is used to determine the dynamic quality benchmark based on the quality assessment results, determine the relative quality difference and optimization priority of the unit based on the quality assessment results and the dynamic quality benchmark, and generate a quality feedback vector. The instruction generation unit is used to determine the copy units to be retained and the copy units to be optimized based on the quality feedback vector, generate optimization actions based on the copy units to be optimized, and generate iterative optimization instructions based on the copy units to be retained, the copy units to be optimized, and the optimization actions.

[0104] Furthermore, the iteration management module includes: The candidate ranking unit is used to calculate the candidate ranking value of the updated candidate in-feed ad copy based on the update quality assessment results, and generate the candidate ranking result based on the candidate ranking value. The degradation verification unit is used to perform non-degradation verification on the candidate information flow ad copy to be verified based on the quality assessment results corresponding to the candidate information flow ad copy in the previous round and the updated quality assessment results corresponding to the candidate information flow ad copy to be verified, and obtain the non-degradation verification result. The iterative decision unit is used to determine whether to continue iterating based on the non-degenerate verification result and the iteration stopping condition, and to determine whether to add the updated candidate information flow ad copy to the output candidate set based on the judgment result.

[0105] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. An iterative optimization method based on feedback from generated content quality assessment, characterized in that, include: Acquire advertising creative task data, which includes product information, core selling points, target audience information, advertising platform information, advertising target information, advertising layout information, brand information, and restriction information; Field identification and constraint extraction are performed on advertising creative task data, and the constraints are categorized to generate creative constraint information; Based on the creative constraint information, a quality assessment configuration and initial candidate information flow ad copy are generated. The quality assessment configuration includes a set of quality assessment indicators and the importance relationship of the indicators. The initial candidate in-feed ad copy was evaluated for copy unit quality, and the quality evaluation results were obtained. A dynamic quality benchmark is determined based on advertising creative task data, creative constraint information, and quality assessment results. A quality feedback vector is generated, an iterative optimization instruction is generated based on the quality feedback vector, and an updated candidate information flow ad copy is generated based on the iterative optimization instruction. The updated candidate information flow ad copy is re-evaluated for quality, ranked and non-degraded, and determined whether to continue iteration. The determination result is obtained, and the iteration instruction is regenerated or the updated candidate information flow ad copy is added to the output candidate set based on the determination result. Determine the target in-feed ad copy from the output candidate set, and output the target in-feed ad copy and the corresponding quality assessment results.

2. The iterative optimization method based on generated content quality assessment feedback according to claim 1, characterized in that, The process of identifying fields and extracting constraints from advertising creative task data, classifying these constraints, and generating creative constraint information specifically includes: Extract product expression constraints and selling point expression constraints based on product information and core selling points; Extract audience expression constraints based on target audience information; Extract platform format constraints based on information from the delivery platform; Extract optimization direction constraints based on the target information; Extract copy structure constraints based on advertisement layout information; Extract brand expression constraints based on brand information; Extract risk boundary constraints based on limiting information; By combining constraints on product expression, selling point expression, audience expression, platform format, optimization direction, copywriting structure, brand expression, and risk boundary, creative constraint information is generated.

3. The iterative optimization method based on generated content quality assessment feedback according to claim 1, characterized in that, The process of generating quality assessment configurations and initial candidate in-feed ad copy based on creative constraint information specifically includes: Content generation control information is generated based on creative constraint information. The content generation control information includes product object, main selling points, target audience, expression style, content structure, distribution platform, title generation requirements, body text generation requirements, content length requirements, and prohibited expression requirements. Multiple initial candidate in-feed ad copy are generated based on content generation control information; A set of quality assessment indicators and their importance relationships are generated based on the target information, platform information, and ad format information.

4. The iterative optimization method based on generated content quality assessment feedback according to claim 1, characterized in that, The process of evaluating the copy unit quality of the initial candidate information flow ad copy to obtain the quality evaluation results specifically includes: Based on the ad layout information and copy structure constraints, the initial candidate information flow ad copy is broken down into multiple copy units, which include a title unit, an opening sentence attraction unit, a pain point expression unit, a product selling point unit, a benefit point expression unit, a call to action unit, and a compliance reminder unit. Based on the set of quality assessment indicators, determine the unit quality performance information of each copywriting unit under the corresponding quality assessment indicators; Calculate the quality score of candidate copywriting based on the unit quality performance information of each copywriting unit; Based on the unit quality performance information and the candidate copy quality scores, a quality assessment result is generated.

5. The iterative optimization method based on generated content quality assessment feedback according to claim 1, characterized in that, The process of determining a dynamic quality benchmark based on advertising creative task data, creative constraint information, and quality assessment results, generating a quality feedback vector, generating iterative optimization instructions based on the quality feedback vector, and generating updated candidate information flow ad copy based on the iterative optimization instructions specifically includes: Determine dynamic quality benchmarks based on quality assessment results; The relative quality differences of units are determined based on the quality assessment results and dynamic quality benchmarks. Determine optimization priorities based on the relative quality differences of the units; Generate a quality feedback vector based on unit quality performance information and optimization priority; The copy units to be retained and the copy units to be optimized are determined based on the quality feedback vector. Generate optimization actions based on the text unit to be optimized; Generate iterative optimization instructions based on the text units to be retained, the text units to be optimized, and the optimization actions; The updated candidate news feed ad copy is generated based on the iterative optimization instructions.

6. The iterative optimization method based on generated content quality assessment feedback according to claim 1, characterized in that, The process of performing quality re-evaluation, candidate ranking, and non-degradation verification on the updated candidate information flow ad copy, determining whether to continue iteration, obtaining a judgment result, and regenerating iteration instructions based on the judgment result or adding the updated candidate information flow ad copy to the output candidate set specifically includes: The quality of the updated candidate in-feed ad copy was re-evaluated to obtain the update quality assessment results. Calculate candidate ranking values ​​based on the updated quality assessment results; Generate candidate sorting results based on candidate sorting values; Perform non-degradation validation on the candidate information flow ad copy to be validated, and obtain the non-degradation validation result; The decision to continue iteration is based on the non-degenerate verification result. If iteration continues, the iteration instruction is regenerated. If iteration does not continue, the updated candidate information flow ad copy is added to the output candidate set.

7. An iterative optimization system based on generated content quality assessment feedback, used to implement the iterative optimization method based on generated content quality assessment feedback as described in any one of claims 1-6, characterized in that, include: The data management module is used to acquire advertising creative task data, identify fields and extract constraints from the advertising creative task data, classify the constraints, and generate creative constraint information. The generation control module is used to generate content generation control information based on creative constraint information, generate initial candidate information flow ad copy based on content generation control information, and generate updated candidate information flow ad copy based on iterative optimization instructions and the previous round of candidate information flow ad copy. The quality assessment module is used to generate quality assessment configuration based on the target information, platform information, and ad format information, and to perform copy unit quality assessment on the initial candidate information flow ad copy or the updated candidate information flow ad copy based on the quality assessment configuration, so as to obtain the quality assessment result or update the quality assessment result. The optimization decision module is used to determine a dynamic quality benchmark based on advertising creative task data, creative constraint information and quality assessment results, generate a quality feedback vector, and generate iterative optimization instructions based on the quality feedback vector; The iteration management module is used to sort candidates and perform non-degradation verification based on the update quality assessment results, and determine whether to continue iteration based on the non-degradation verification results. If iteration continues, the iteration instruction is regenerated; if iteration does not continue, the updated candidate information flow ad copy is added to the output candidate set. The output management module is used to determine the target information flow ad copy from the set of output candidates and output the target information flow ad copy and the corresponding quality assessment results.

8. The iterative optimization system based on generated content quality assessment feedback according to claim 7, characterized in that, The data management module includes: The data recognition unit is used to acquire and identify product information, core selling points, target audience information, advertising platform information, advertising target information, advertising layout information, brand information, and restriction information in advertising creative task data; The constraint processing unit is used to extract product expression constraints, selling point expression constraints, audience expression constraints, platform format constraints, optimization direction constraints, copywriting structure constraints, brand expression constraints, and risk boundary constraints from advertising creative task data, and to classify each constraint to generate creative constraint information; The generation control module includes: A configuration unit is generated to generate content generation control information based on creative constraint information. The candidate generation unit is used to generate initial candidate information flow ad copy based on content generation control information, and to generate updated candidate information flow ad copy based on iterative optimization instructions, content generation control information, and the previous round of candidate information flow ad copy.

9. The iterative optimization system based on generated content quality assessment feedback according to claim 7, characterized in that, The quality assessment module includes: The evaluation configuration unit is used to generate a set of quality evaluation indicators and the importance relationships of the indicators based on the target information, the platform information, and the ad format information. The unit evaluation unit is used to break down candidate infeed ad copy into multiple copy units, determine the unit quality performance information of each copy unit based on the quality assessment indicator set and the importance relationship of the indicators, calculate the candidate copy quality score, and generate quality assessment results. The optimization decision module includes: The quality analysis unit is used to determine the dynamic quality benchmark based on the quality assessment results, determine the relative quality difference and optimization priority of the unit based on the quality assessment results and the dynamic quality benchmark, and generate a quality feedback vector. The instruction generation unit is used to determine the copy units to be retained and the copy units to be optimized based on the quality feedback vector, generate optimization actions based on the copy units to be optimized, and generate iterative optimization instructions based on the copy units to be retained, the copy units to be optimized, and the optimization actions.

10. The iterative optimization system based on generated content quality assessment feedback according to claim 7, characterized in that, The iteration management module includes: The candidate ranking unit is used to calculate the candidate ranking value of the updated candidate in-feed ad copy based on the update quality assessment results, and generate the candidate ranking result based on the candidate ranking value. The degradation verification unit is used to perform non-degradation verification on the candidate information flow ad copy to be verified based on the quality assessment results corresponding to the candidate information flow ad copy in the previous round and the updated quality assessment results corresponding to the candidate information flow ad copy to be verified, and obtain the non-degradation verification result. The iterative decision unit is used to determine whether to continue iterating based on the non-degenerate verification result and the iteration stopping condition, and to determine whether to add the updated candidate information flow ad copy to the output candidate set based on the judgment result.

Images